zig/src/link/Wasm.zig
Andrew Kelley 3fc6fc6812 std.builtin.Endian: make the tags lower case
Let's take this breaking change opportunity to fix the style of this
enum.
2023-10-31 21:37:35 -04:00

5021 lines
205 KiB
Zig

const Wasm = @This();
const std = @import("std");
const builtin = @import("builtin");
const mem = std.mem;
const Allocator = std.mem.Allocator;
const assert = std.debug.assert;
const fs = std.fs;
const leb = std.leb;
const log = std.log.scoped(.link);
pub const Atom = @import("Wasm/Atom.zig");
const Dwarf = @import("Dwarf.zig");
const Module = @import("../Module.zig");
const InternPool = @import("../InternPool.zig");
const Compilation = @import("../Compilation.zig");
const CodeGen = @import("../arch/wasm/CodeGen.zig");
const codegen = @import("../codegen.zig");
const link = @import("../link.zig");
const lldMain = @import("../main.zig").lldMain;
const trace = @import("../tracy.zig").trace;
const build_options = @import("build_options");
const wasi_libc = @import("../wasi_libc.zig");
const Cache = std.Build.Cache;
const Type = @import("../type.zig").Type;
const TypedValue = @import("../TypedValue.zig");
const LlvmObject = @import("../codegen/llvm.zig").Object;
const Air = @import("../Air.zig");
const Liveness = @import("../Liveness.zig");
const Symbol = @import("Wasm/Symbol.zig");
const Object = @import("Wasm/Object.zig");
const Archive = @import("Wasm/Archive.zig");
const types = @import("Wasm/types.zig");
pub const Relocation = types.Relocation;
pub const base_tag: link.File.Tag = .wasm;
base: link.File,
/// Output name of the file
name: []const u8,
/// If this is not null, an object file is created by LLVM and linked with LLD afterwards.
llvm_object: ?*LlvmObject = null,
/// When importing objects from the host environment, a name must be supplied.
/// LLVM uses "env" by default when none is given. This would be a good default for Zig
/// to support existing code.
/// TODO: Allow setting this through a flag?
host_name: []const u8 = "env",
/// List of all `Decl` that are currently alive.
/// Each index maps to the corresponding `Atom.Index`.
decls: std.AutoHashMapUnmanaged(Module.Decl.Index, Atom.Index) = .{},
/// Mapping between an `Atom` and its type index representing the Wasm
/// type of the function signature.
atom_types: std.AutoHashMapUnmanaged(Atom.Index, u32) = .{},
/// List of all symbols generated by Zig code.
symbols: std.ArrayListUnmanaged(Symbol) = .{},
/// List of symbol indexes which are free to be used.
symbols_free_list: std.ArrayListUnmanaged(u32) = .{},
/// Maps atoms to their segment index
atoms: std.AutoHashMapUnmanaged(u32, Atom.Index) = .{},
/// List of all atoms.
managed_atoms: std.ArrayListUnmanaged(Atom) = .{},
/// Represents the index into `segments` where the 'code' section
/// lives.
code_section_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_info' section.
debug_info_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_line' section.
debug_line_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_loc' section.
debug_loc_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_ranges' section.
debug_ranges_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_pubnames' section.
debug_pubnames_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_pubtypes' section.
debug_pubtypes_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_pubtypes' section.
debug_str_index: ?u32 = null,
/// The index of the segment representing the custom '.debug_pubtypes' section.
debug_abbrev_index: ?u32 = null,
/// The count of imported functions. This number will be appended
/// to the function indexes as their index starts at the lowest non-extern function.
imported_functions_count: u32 = 0,
/// The count of imported wasm globals. This number will be appended
/// to the global indexes when sections are merged.
imported_globals_count: u32 = 0,
/// The count of imported tables. This number will be appended
/// to the table indexes when sections are merged.
imported_tables_count: u32 = 0,
/// Map of symbol locations, represented by its `types.Import`
imports: std.AutoHashMapUnmanaged(SymbolLoc, types.Import) = .{},
/// Represents non-synthetic section entries.
/// Used for code, data and custom sections.
segments: std.ArrayListUnmanaged(Segment) = .{},
/// Maps a data segment key (such as .rodata) to the index into `segments`.
data_segments: std.StringArrayHashMapUnmanaged(u32) = .{},
/// A table of `types.Segment` which provide meta data
/// about a data symbol such as its name where the key is
/// the segment index, which can be found from `data_segments`
segment_info: std.AutoArrayHashMapUnmanaged(u32, types.Segment) = .{},
/// Deduplicated string table for strings used by symbols, imports and exports.
string_table: StringTable = .{},
/// Debug information for wasm
dwarf: ?Dwarf = null,
// Output sections
/// Output type section
func_types: std.ArrayListUnmanaged(std.wasm.Type) = .{},
/// Output function section where the key is the original
/// function index and the value is function.
/// This allows us to map multiple symbols to the same function.
functions: std.AutoArrayHashMapUnmanaged(struct { file: ?u16, index: u32 }, std.wasm.Func) = .{},
/// Output global section
wasm_globals: std.ArrayListUnmanaged(std.wasm.Global) = .{},
/// Memory section
memories: std.wasm.Memory = .{ .limits = .{
.min = 0,
.max = undefined,
.flags = 0,
} },
/// Output table section
tables: std.ArrayListUnmanaged(std.wasm.Table) = .{},
/// Output export section
exports: std.ArrayListUnmanaged(types.Export) = .{},
/// List of initialization functions. These must be called in order of priority
/// by the (synthetic) __wasm_call_ctors function.
init_funcs: std.ArrayListUnmanaged(InitFuncLoc) = .{},
/// Index to a function defining the entry of the wasm file
entry: ?u32 = null,
/// Indirect function table, used to call function pointers
/// When this is non-zero, we must emit a table entry,
/// as well as an 'elements' section.
///
/// Note: Key is symbol location, value represents the index into the table
function_table: std.AutoHashMapUnmanaged(SymbolLoc, u32) = .{},
/// All object files and their data which are linked into the final binary
objects: std.ArrayListUnmanaged(Object) = .{},
/// All archive files that are lazy loaded.
/// e.g. when an undefined symbol references a symbol from the archive.
archives: std.ArrayListUnmanaged(Archive) = .{},
/// A map of global names (read: offset into string table) to their symbol location
globals: std.AutoHashMapUnmanaged(u32, SymbolLoc) = .{},
/// The list of GOT symbols and their location
got_symbols: std.ArrayListUnmanaged(SymbolLoc) = .{},
/// Maps discarded symbols and their positions to the location of the symbol
/// it was resolved to
discarded: std.AutoHashMapUnmanaged(SymbolLoc, SymbolLoc) = .{},
/// List of all symbol locations which have been resolved by the linker and will be emit
/// into the final binary.
resolved_symbols: std.AutoArrayHashMapUnmanaged(SymbolLoc, void) = .{},
/// Symbols that remain undefined after symbol resolution.
/// Note: The key represents an offset into the string table, rather than the actual string.
undefs: std.AutoArrayHashMapUnmanaged(u32, SymbolLoc) = .{},
/// Maps a symbol's location to an atom. This can be used to find meta
/// data of a symbol, such as its size, or its offset to perform a relocation.
/// Undefined (and synthetic) symbols do not have an Atom and therefore cannot be mapped.
symbol_atom: std.AutoHashMapUnmanaged(SymbolLoc, Atom.Index) = .{},
/// Maps a symbol's location to its export name, which may differ from the decl's name
/// which does the exporting.
/// Note: The value represents the offset into the string table, rather than the actual string.
export_names: std.AutoHashMapUnmanaged(SymbolLoc, u32) = .{},
/// Represents the symbol index of the error name table
/// When this is `null`, no code references an error using runtime `@errorName`.
/// During initializion, a symbol with corresponding atom will be created that is
/// used to perform relocations to the pointer of this table.
/// The actual table is populated during `flush`.
error_table_symbol: ?u32 = null,
// Debug section atoms. These are only set when the current compilation
// unit contains Zig code. The lifetime of these atoms are extended
// until the end of the compiler's lifetime. Meaning they're not freed
// during `flush()` in incremental-mode.
debug_info_atom: ?Atom.Index = null,
debug_line_atom: ?Atom.Index = null,
debug_loc_atom: ?Atom.Index = null,
debug_ranges_atom: ?Atom.Index = null,
debug_abbrev_atom: ?Atom.Index = null,
debug_str_atom: ?Atom.Index = null,
debug_pubnames_atom: ?Atom.Index = null,
debug_pubtypes_atom: ?Atom.Index = null,
/// List of atom indexes of functions that are generated by the backend,
/// rather than by the linker.
synthetic_functions: std.ArrayListUnmanaged(Atom.Index) = .{},
/// Map for storing anonymous declarations. Each anonymous decl maps to its Atom's index.
anon_decls: std.AutoArrayHashMapUnmanaged(InternPool.Index, Atom.Index) = .{},
pub const Alignment = types.Alignment;
pub const Segment = struct {
alignment: Alignment,
size: u32,
offset: u32,
flags: u32,
pub const Flag = enum(u32) {
WASM_DATA_SEGMENT_IS_PASSIVE = 0x01,
WASM_DATA_SEGMENT_HAS_MEMINDEX = 0x02,
};
pub fn isPassive(segment: Segment) bool {
return segment.flags & @intFromEnum(Flag.WASM_DATA_SEGMENT_IS_PASSIVE) != 0;
}
/// For a given segment, determines if it needs passive initialization
fn needsPassiveInitialization(segment: Segment, import_mem: bool, name: []const u8) bool {
if (import_mem and !std.mem.eql(u8, name, ".bss")) {
return true;
}
return segment.isPassive();
}
};
pub const Export = struct {
sym_index: ?u32 = null,
};
pub const SymbolLoc = struct {
/// The index of the symbol within the specified file
index: u32,
/// The index of the object file where the symbol resides.
/// When this is `null` the symbol comes from a non-object file.
file: ?u16,
/// From a given location, returns the corresponding symbol in the wasm binary
pub fn getSymbol(loc: SymbolLoc, wasm_bin: *const Wasm) *Symbol {
if (wasm_bin.discarded.get(loc)) |new_loc| {
return new_loc.getSymbol(wasm_bin);
}
if (loc.file) |object_index| {
const object = wasm_bin.objects.items[object_index];
return &object.symtable[loc.index];
}
return &wasm_bin.symbols.items[loc.index];
}
/// From a given location, returns the name of the symbol.
pub fn getName(loc: SymbolLoc, wasm_bin: *const Wasm) []const u8 {
if (wasm_bin.discarded.get(loc)) |new_loc| {
return new_loc.getName(wasm_bin);
}
if (loc.file) |object_index| {
const object = wasm_bin.objects.items[object_index];
return object.string_table.get(object.symtable[loc.index].name);
}
return wasm_bin.string_table.get(wasm_bin.symbols.items[loc.index].name);
}
/// From a given symbol location, returns the final location.
/// e.g. when a symbol was resolved and replaced by the symbol
/// in a different file, this will return said location.
/// If the symbol wasn't replaced by another, this will return
/// the given location itwasm.
pub fn finalLoc(loc: SymbolLoc, wasm_bin: *const Wasm) SymbolLoc {
if (wasm_bin.discarded.get(loc)) |new_loc| {
return new_loc.finalLoc(wasm_bin);
}
return loc;
}
};
// Contains the location of the function symbol, as well as
/// the priority itself of the initialization function.
pub const InitFuncLoc = struct {
/// object file index in the list of objects.
/// Unlike `SymbolLoc` this cannot be `null` as we never define
/// our own ctors.
file: u16,
/// Symbol index within the corresponding object file.
index: u32,
/// The priority in which the constructor must be called.
priority: u32,
/// From a given `InitFuncLoc` returns the corresponding function symbol
fn getSymbol(loc: InitFuncLoc, wasm: *const Wasm) *Symbol {
return getSymbolLoc(loc).getSymbol(wasm);
}
/// Turns the given `InitFuncLoc` into a `SymbolLoc`
fn getSymbolLoc(loc: InitFuncLoc) SymbolLoc {
return .{ .file = loc.file, .index = loc.index };
}
/// Returns true when `lhs` has a higher priority (e.i. value closer to 0) than `rhs`.
fn lessThan(ctx: void, lhs: InitFuncLoc, rhs: InitFuncLoc) bool {
_ = ctx;
return lhs.priority < rhs.priority;
}
};
/// Generic string table that duplicates strings
/// and converts them into offsets instead.
pub const StringTable = struct {
/// Table that maps string offsets, which is used to de-duplicate strings.
/// Rather than having the offset map to the data, the `StringContext` holds all bytes of the string.
/// The strings are stored as a contigious array where each string is zero-terminated.
string_table: std.HashMapUnmanaged(
u32,
void,
std.hash_map.StringIndexContext,
std.hash_map.default_max_load_percentage,
) = .{},
/// Holds the actual data of the string table.
string_data: std.ArrayListUnmanaged(u8) = .{},
/// Accepts a string and searches for a corresponding string.
/// When found, de-duplicates the string and returns the existing offset instead.
/// When the string is not found in the `string_table`, a new entry will be inserted
/// and the new offset to its data will be returned.
pub fn put(table: *StringTable, allocator: Allocator, string: []const u8) !u32 {
const gop = try table.string_table.getOrPutContextAdapted(
allocator,
string,
std.hash_map.StringIndexAdapter{ .bytes = &table.string_data },
.{ .bytes = &table.string_data },
);
if (gop.found_existing) {
const off = gop.key_ptr.*;
log.debug("reusing string '{s}' at offset 0x{x}", .{ string, off });
return off;
}
try table.string_data.ensureUnusedCapacity(allocator, string.len + 1);
const offset = @as(u32, @intCast(table.string_data.items.len));
log.debug("writing new string '{s}' at offset 0x{x}", .{ string, offset });
table.string_data.appendSliceAssumeCapacity(string);
table.string_data.appendAssumeCapacity(0);
gop.key_ptr.* = offset;
return offset;
}
/// From a given offset, returns its corresponding string value.
/// Asserts offset does not exceed bounds.
pub fn get(table: StringTable, off: u32) []const u8 {
assert(off < table.string_data.items.len);
return mem.sliceTo(@as([*:0]const u8, @ptrCast(table.string_data.items.ptr + off)), 0);
}
/// Returns the offset of a given string when it exists.
/// Will return null if the given string does not yet exist within the string table.
pub fn getOffset(table: *StringTable, string: []const u8) ?u32 {
return table.string_table.getKeyAdapted(
string,
std.hash_map.StringIndexAdapter{ .bytes = &table.string_data },
);
}
/// Frees all resources of the string table. Any references pointing
/// to the strings will be invalid.
pub fn deinit(table: *StringTable, allocator: Allocator) void {
table.string_data.deinit(allocator);
table.string_table.deinit(allocator);
table.* = undefined;
}
};
pub fn openPath(allocator: Allocator, sub_path: []const u8, options: link.Options) !*Wasm {
assert(options.target.ofmt == .wasm);
if (options.use_llvm and options.use_lld) {
return createEmpty(allocator, options);
}
const wasm_bin = try createEmpty(allocator, options);
errdefer wasm_bin.base.destroy();
// We are not using LLD at this point, so ensure we set the intermediary basename
if (build_options.have_llvm and options.use_llvm and options.module != null) {
// TODO this intermediary_basename isn't enough; in the case of `zig build-exe`,
// we also want to put the intermediary object file in the cache while the
// main emit directory is the cwd.
wasm_bin.base.intermediary_basename = try std.fmt.allocPrint(allocator, "{s}{s}", .{
options.emit.?.sub_path, options.target.ofmt.fileExt(options.target.cpu.arch),
});
}
// TODO: read the file and keep valid parts instead of truncating
const file = try options.emit.?.directory.handle.createFile(sub_path, .{
.truncate = true,
.read = true,
.mode = if (fs.has_executable_bit)
if (options.target.os.tag == .wasi and options.output_mode == .Exe)
fs.File.default_mode | 0b001_000_000
else
fs.File.default_mode
else
0,
});
wasm_bin.base.file = file;
wasm_bin.name = sub_path;
// create stack pointer symbol
{
const loc = try wasm_bin.createSyntheticSymbol("__stack_pointer", .global);
const symbol = loc.getSymbol(wasm_bin);
// For object files we will import the stack pointer symbol
if (options.output_mode == .Obj) {
symbol.setUndefined(true);
symbol.index = @as(u32, @intCast(wasm_bin.imported_globals_count));
wasm_bin.imported_globals_count += 1;
try wasm_bin.imports.putNoClobber(
allocator,
loc,
.{
.module_name = try wasm_bin.string_table.put(allocator, wasm_bin.host_name),
.name = symbol.name,
.kind = .{ .global = .{ .valtype = .i32, .mutable = true } },
},
);
} else {
symbol.index = @intCast(wasm_bin.imported_globals_count + wasm_bin.wasm_globals.items.len);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
const global = try wasm_bin.wasm_globals.addOne(allocator);
global.* = .{
.global_type = .{
.valtype = .i32,
.mutable = true,
},
.init = .{ .i32_const = 0 },
};
}
}
// create indirect function pointer symbol
{
const loc = try wasm_bin.createSyntheticSymbol("__indirect_function_table", .table);
const symbol = loc.getSymbol(wasm_bin);
const table: std.wasm.Table = .{
.limits = .{ .flags = 0, .min = 0, .max = undefined }, // will be overwritten during `mapFunctionTable`
.reftype = .funcref,
};
if (options.output_mode == .Obj or options.import_table) {
symbol.setUndefined(true);
symbol.index = @intCast(wasm_bin.imported_tables_count);
wasm_bin.imported_tables_count += 1;
try wasm_bin.imports.put(allocator, loc, .{
.module_name = try wasm_bin.string_table.put(allocator, wasm_bin.host_name),
.name = symbol.name,
.kind = .{ .table = table },
});
} else {
symbol.index = @as(u32, @intCast(wasm_bin.imported_tables_count + wasm_bin.tables.items.len));
try wasm_bin.tables.append(allocator, table);
if (options.export_table) {
symbol.setFlag(.WASM_SYM_EXPORTED);
} else {
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
}
}
}
// create __wasm_call_ctors
{
const loc = try wasm_bin.createSyntheticSymbol("__wasm_call_ctors", .function);
const symbol = loc.getSymbol(wasm_bin);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
// we do not know the function index until after we merged all sections.
// Therefore we set `symbol.index` and create its corresponding references
// at the end during `initializeCallCtorsFunction`.
}
// shared-memory symbols for TLS support
if (wasm_bin.base.options.shared_memory) {
{
const loc = try wasm_bin.createSyntheticSymbol("__tls_base", .global);
const symbol = loc.getSymbol(wasm_bin);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
symbol.index = @intCast(wasm_bin.imported_globals_count + wasm_bin.wasm_globals.items.len);
try wasm_bin.wasm_globals.append(wasm_bin.base.allocator, .{
.global_type = .{ .valtype = .i32, .mutable = true },
.init = .{ .i32_const = undefined },
});
}
{
const loc = try wasm_bin.createSyntheticSymbol("__tls_size", .global);
const symbol = loc.getSymbol(wasm_bin);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
symbol.index = @intCast(wasm_bin.imported_globals_count + wasm_bin.wasm_globals.items.len);
try wasm_bin.wasm_globals.append(wasm_bin.base.allocator, .{
.global_type = .{ .valtype = .i32, .mutable = false },
.init = .{ .i32_const = undefined },
});
}
{
const loc = try wasm_bin.createSyntheticSymbol("__tls_align", .global);
const symbol = loc.getSymbol(wasm_bin);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
symbol.index = @intCast(wasm_bin.imported_globals_count + wasm_bin.wasm_globals.items.len);
try wasm_bin.wasm_globals.append(wasm_bin.base.allocator, .{
.global_type = .{ .valtype = .i32, .mutable = false },
.init = .{ .i32_const = undefined },
});
}
{
const loc = try wasm_bin.createSyntheticSymbol("__wasm_init_tls", .function);
const symbol = loc.getSymbol(wasm_bin);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
}
}
// if (!options.strip and options.module != null) {
// wasm_bin.dwarf = Dwarf.init(allocator, &wasm_bin.base, .dwarf32);
// try wasm_bin.initDebugSections();
// }
return wasm_bin;
}
pub fn createEmpty(gpa: Allocator, options: link.Options) !*Wasm {
const wasm = try gpa.create(Wasm);
errdefer gpa.destroy(wasm);
wasm.* = .{
.base = .{
.tag = .wasm,
.options = options,
.file = null,
.allocator = gpa,
},
.name = undefined,
};
if (options.use_llvm) {
wasm.llvm_object = try LlvmObject.create(gpa, options);
}
return wasm;
}
/// For a given name, creates a new global synthetic symbol.
/// Leaves index undefined and the default flags (0).
fn createSyntheticSymbol(wasm: *Wasm, name: []const u8, tag: Symbol.Tag) !SymbolLoc {
const name_offset = try wasm.string_table.put(wasm.base.allocator, name);
return wasm.createSyntheticSymbolOffset(name_offset, tag);
}
fn createSyntheticSymbolOffset(wasm: *Wasm, name_offset: u32, tag: Symbol.Tag) !SymbolLoc {
const sym_index = @as(u32, @intCast(wasm.symbols.items.len));
const loc: SymbolLoc = .{ .index = sym_index, .file = null };
try wasm.symbols.append(wasm.base.allocator, .{
.name = name_offset,
.flags = 0,
.tag = tag,
.index = undefined,
.virtual_address = undefined,
});
try wasm.resolved_symbols.putNoClobber(wasm.base.allocator, loc, {});
try wasm.globals.put(wasm.base.allocator, name_offset, loc);
return loc;
}
/// Initializes symbols and atoms for the debug sections
/// Initialization is only done when compiling Zig code.
/// When Zig is invoked as a linker instead, the atoms
/// and symbols come from the object files instead.
pub fn initDebugSections(wasm: *Wasm) !void {
if (wasm.dwarf == null) return; // not compiling Zig code, so no need to pre-initialize debug sections
assert(wasm.debug_info_index == null);
// this will create an Atom and set the index for us.
wasm.debug_info_atom = try wasm.createDebugSectionForIndex(&wasm.debug_info_index, ".debug_info");
wasm.debug_line_atom = try wasm.createDebugSectionForIndex(&wasm.debug_line_index, ".debug_line");
wasm.debug_loc_atom = try wasm.createDebugSectionForIndex(&wasm.debug_loc_index, ".debug_loc");
wasm.debug_abbrev_atom = try wasm.createDebugSectionForIndex(&wasm.debug_abbrev_index, ".debug_abbrev");
wasm.debug_ranges_atom = try wasm.createDebugSectionForIndex(&wasm.debug_ranges_index, ".debug_ranges");
wasm.debug_str_atom = try wasm.createDebugSectionForIndex(&wasm.debug_str_index, ".debug_str");
wasm.debug_pubnames_atom = try wasm.createDebugSectionForIndex(&wasm.debug_pubnames_index, ".debug_pubnames");
wasm.debug_pubtypes_atom = try wasm.createDebugSectionForIndex(&wasm.debug_pubtypes_index, ".debug_pubtypes");
}
fn parseInputFiles(wasm: *Wasm, files: []const []const u8) !void {
for (files) |path| {
if (try wasm.parseObjectFile(path)) continue;
if (try wasm.parseArchive(path, false)) continue; // load archives lazily
log.warn("Unexpected file format at path: '{s}'", .{path});
}
}
/// Parses the object file from given path. Returns true when the given file was an object
/// file and parsed successfully. Returns false when file is not an object file.
/// May return an error instead when parsing failed.
fn parseObjectFile(wasm: *Wasm, path: []const u8) !bool {
const file = try fs.cwd().openFile(path, .{});
errdefer file.close();
var object = Object.create(wasm.base.allocator, file, path, null) catch |err| switch (err) {
error.InvalidMagicByte, error.NotObjectFile => return false,
else => |e| return e,
};
errdefer object.deinit(wasm.base.allocator);
try wasm.objects.append(wasm.base.allocator, object);
return true;
}
/// For a given `Module.Decl.Index` returns its corresponding `Atom.Index`.
/// When the index was not found, a new `Atom` will be created, and its index will be returned.
/// The newly created Atom is empty with default fields as specified by `Atom.empty`.
pub fn getOrCreateAtomForDecl(wasm: *Wasm, decl_index: Module.Decl.Index) !Atom.Index {
const gop = try wasm.decls.getOrPut(wasm.base.allocator, decl_index);
if (!gop.found_existing) {
gop.value_ptr.* = try wasm.createAtom();
}
return gop.value_ptr.*;
}
/// Creates a new empty `Atom` and returns its `Atom.Index`
fn createAtom(wasm: *Wasm) !Atom.Index {
const index = @as(Atom.Index, @intCast(wasm.managed_atoms.items.len));
const atom = try wasm.managed_atoms.addOne(wasm.base.allocator);
atom.* = Atom.empty;
atom.sym_index = try wasm.allocateSymbol();
try wasm.symbol_atom.putNoClobber(wasm.base.allocator, .{ .file = null, .index = atom.sym_index }, index);
return index;
}
pub inline fn getAtom(wasm: *const Wasm, index: Atom.Index) Atom {
return wasm.managed_atoms.items[index];
}
pub inline fn getAtomPtr(wasm: *Wasm, index: Atom.Index) *Atom {
return &wasm.managed_atoms.items[index];
}
/// Parses an archive file and will then parse each object file
/// that was found in the archive file.
/// Returns false when the file is not an archive file.
/// May return an error instead when parsing failed.
///
/// When `force_load` is `true`, it will for link all object files in the archive.
/// When false, it will only link with object files that contain symbols that
/// are referenced by other object files or Zig code.
fn parseArchive(wasm: *Wasm, path: []const u8, force_load: bool) !bool {
const file = try fs.cwd().openFile(path, .{});
errdefer file.close();
var archive: Archive = .{
.file = file,
.name = path,
};
archive.parse(wasm.base.allocator) catch |err| switch (err) {
error.EndOfStream, error.NotArchive => {
archive.deinit(wasm.base.allocator);
return false;
},
else => |e| return e,
};
if (!force_load) {
errdefer archive.deinit(wasm.base.allocator);
try wasm.archives.append(wasm.base.allocator, archive);
return true;
}
defer archive.deinit(wasm.base.allocator);
// In this case we must force link all embedded object files within the archive
// We loop over all symbols, and then group them by offset as the offset
// notates where the object file starts.
var offsets = std.AutoArrayHashMap(u32, void).init(wasm.base.allocator);
defer offsets.deinit();
for (archive.toc.values()) |symbol_offsets| {
for (symbol_offsets.items) |sym_offset| {
try offsets.put(sym_offset, {});
}
}
for (offsets.keys()) |file_offset| {
const object = try wasm.objects.addOne(wasm.base.allocator);
object.* = try archive.parseObject(wasm.base.allocator, file_offset);
}
return true;
}
fn requiresTLSReloc(wasm: *const Wasm) bool {
for (wasm.got_symbols.items) |loc| {
if (loc.getSymbol(wasm).isTLS()) {
return true;
}
}
return false;
}
fn resolveSymbolsInObject(wasm: *Wasm, object_index: u16) !void {
const object: Object = wasm.objects.items[object_index];
log.debug("Resolving symbols in object: '{s}'", .{object.name});
for (object.symtable, 0..) |symbol, i| {
const sym_index = @as(u32, @intCast(i));
const location: SymbolLoc = .{
.file = object_index,
.index = sym_index,
};
const sym_name = object.string_table.get(symbol.name);
if (mem.eql(u8, sym_name, "__indirect_function_table")) {
continue;
}
const sym_name_index = try wasm.string_table.put(wasm.base.allocator, sym_name);
if (symbol.isLocal()) {
if (symbol.isUndefined()) {
log.err("Local symbols are not allowed to reference imports", .{});
log.err(" symbol '{s}' defined in '{s}'", .{ sym_name, object.name });
return error.UndefinedLocal;
}
try wasm.resolved_symbols.putNoClobber(wasm.base.allocator, location, {});
continue;
}
const maybe_existing = try wasm.globals.getOrPut(wasm.base.allocator, sym_name_index);
if (!maybe_existing.found_existing) {
maybe_existing.value_ptr.* = location;
try wasm.resolved_symbols.putNoClobber(wasm.base.allocator, location, {});
if (symbol.isUndefined()) {
try wasm.undefs.putNoClobber(wasm.base.allocator, sym_name_index, location);
}
continue;
}
const existing_loc = maybe_existing.value_ptr.*;
const existing_sym: *Symbol = existing_loc.getSymbol(wasm);
const existing_file_path = if (existing_loc.file) |file| blk: {
break :blk wasm.objects.items[file].name;
} else wasm.name;
if (!existing_sym.isUndefined()) outer: {
if (!symbol.isUndefined()) inner: {
if (symbol.isWeak()) {
break :inner; // ignore the new symbol (discard it)
}
if (existing_sym.isWeak()) {
break :outer; // existing is weak, while new one isn't. Replace it.
}
// both are defined and weak, we have a symbol collision.
log.err("symbol '{s}' defined multiple times", .{sym_name});
log.err(" first definition in '{s}'", .{existing_file_path});
log.err(" next definition in '{s}'", .{object.name});
return error.SymbolCollision;
}
try wasm.discarded.put(wasm.base.allocator, location, existing_loc);
continue; // Do not overwrite defined symbols with undefined symbols
}
if (symbol.tag != existing_sym.tag) {
log.err("symbol '{s}' mismatching type '{s}", .{ sym_name, @tagName(symbol.tag) });
log.err(" first definition in '{s}'", .{existing_file_path});
log.err(" next definition in '{s}'", .{object.name});
return error.SymbolMismatchingType;
}
if (existing_sym.isUndefined() and symbol.isUndefined()) {
// only verify module/import name for function symbols
if (symbol.tag == .function) {
const existing_name = if (existing_loc.file) |file_index| blk: {
const obj = wasm.objects.items[file_index];
const name_index = obj.findImport(symbol.tag.externalType(), existing_sym.index).module_name;
break :blk obj.string_table.get(name_index);
} else blk: {
const name_index = wasm.imports.get(existing_loc).?.module_name;
break :blk wasm.string_table.get(name_index);
};
const module_index = object.findImport(symbol.tag.externalType(), symbol.index).module_name;
const module_name = object.string_table.get(module_index);
if (!mem.eql(u8, existing_name, module_name)) {
log.err("symbol '{s}' module name mismatch. Expected '{s}', but found '{s}'", .{
sym_name,
existing_name,
module_name,
});
log.err(" first definition in '{s}'", .{existing_file_path});
log.err(" next definition in '{s}'", .{object.name});
return error.ModuleNameMismatch;
}
}
// both undefined so skip overwriting existing symbol and discard the new symbol
try wasm.discarded.put(wasm.base.allocator, location, existing_loc);
continue;
}
if (existing_sym.tag == .global) {
const existing_ty = wasm.getGlobalType(existing_loc);
const new_ty = wasm.getGlobalType(location);
if (existing_ty.mutable != new_ty.mutable or existing_ty.valtype != new_ty.valtype) {
log.err("symbol '{s}' mismatching global types", .{sym_name});
log.err(" first definition in '{s}'", .{existing_file_path});
log.err(" next definition in '{s}'", .{object.name});
return error.GlobalTypeMismatch;
}
}
if (existing_sym.tag == .function) {
const existing_ty = wasm.getFunctionSignature(existing_loc);
const new_ty = wasm.getFunctionSignature(location);
if (!existing_ty.eql(new_ty)) {
log.err("symbol '{s}' mismatching function signatures.", .{sym_name});
log.err(" expected signature {}, but found signature {}", .{ existing_ty, new_ty });
log.err(" first definition in '{s}'", .{existing_file_path});
log.err(" next definition in '{s}'", .{object.name});
return error.FunctionSignatureMismatch;
}
}
// when both symbols are weak, we skip overwriting unless the existing
// symbol is weak and the new one isn't, in which case we *do* overwrite it.
if (existing_sym.isWeak() and symbol.isWeak()) blk: {
if (existing_sym.isUndefined() and !symbol.isUndefined()) break :blk;
try wasm.discarded.put(wasm.base.allocator, location, existing_loc);
continue;
}
// simply overwrite with the new symbol
log.debug("Overwriting symbol '{s}'", .{sym_name});
log.debug(" old definition in '{s}'", .{existing_file_path});
log.debug(" new definition in '{s}'", .{object.name});
try wasm.discarded.putNoClobber(wasm.base.allocator, existing_loc, location);
maybe_existing.value_ptr.* = location;
try wasm.globals.put(wasm.base.allocator, sym_name_index, location);
try wasm.resolved_symbols.put(wasm.base.allocator, location, {});
assert(wasm.resolved_symbols.swapRemove(existing_loc));
if (existing_sym.isUndefined()) {
_ = wasm.undefs.swapRemove(sym_name_index);
}
}
}
fn resolveSymbolsInArchives(wasm: *Wasm) !void {
if (wasm.archives.items.len == 0) return;
log.debug("Resolving symbols in archives", .{});
var index: u32 = 0;
undef_loop: while (index < wasm.undefs.count()) {
const sym_name_index = wasm.undefs.keys()[index];
for (wasm.archives.items) |archive| {
const sym_name = wasm.string_table.get(sym_name_index);
log.debug("Detected symbol '{s}' in archive '{s}', parsing objects..", .{ sym_name, archive.name });
const offset = archive.toc.get(sym_name) orelse {
// symbol does not exist in this archive
continue;
};
// Symbol is found in unparsed object file within current archive.
// Parse object and and resolve symbols again before we check remaining
// undefined symbols.
const object_file_index = @as(u16, @intCast(wasm.objects.items.len));
var object = try archive.parseObject(wasm.base.allocator, offset.items[0]);
try wasm.objects.append(wasm.base.allocator, object);
try wasm.resolveSymbolsInObject(object_file_index);
// continue loop for any remaining undefined symbols that still exist
// after resolving last object file
continue :undef_loop;
}
index += 1;
}
}
/// Writes an unsigned 32-bit integer as a LEB128-encoded 'i32.const' value.
fn writeI32Const(writer: anytype, val: u32) !void {
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeILEB128(writer, @as(i32, @bitCast(val)));
}
fn setupInitMemoryFunction(wasm: *Wasm) !void {
// Passive segments are used to avoid memory being reinitialized on each
// thread's instantiation. These passive segments are initialized and
// dropped in __wasm_init_memory, which is registered as the start function
// We also initialize bss segments (using memory.fill) as part of this
// function.
if (!wasm.hasPassiveInitializationSegments()) {
return;
}
const flag_address: u32 = if (wasm.base.options.shared_memory) address: {
// when we have passive initialization segments and shared memory
// `setupMemory` will create this symbol and set its virtual address.
const loc = wasm.findGlobalSymbol("__wasm_init_memory_flag").?;
break :address loc.getSymbol(wasm).virtual_address;
} else 0;
var function_body = std.ArrayList(u8).init(wasm.base.allocator);
defer function_body.deinit();
const writer = function_body.writer();
// we have 0 locals
try leb.writeULEB128(writer, @as(u32, 0));
if (wasm.base.options.shared_memory) {
// destination blocks
// based on values we jump to corresponding label
try writer.writeByte(std.wasm.opcode(.block)); // $drop
try writer.writeByte(std.wasm.block_empty); // block type
try writer.writeByte(std.wasm.opcode(.block)); // $wait
try writer.writeByte(std.wasm.block_empty); // block type
try writer.writeByte(std.wasm.opcode(.block)); // $init
try writer.writeByte(std.wasm.block_empty); // block type
// atomically check
try writeI32Const(writer, flag_address);
try writeI32Const(writer, 0);
try writeI32Const(writer, 1);
try writer.writeByte(std.wasm.opcode(.atomics_prefix));
try leb.writeULEB128(writer, std.wasm.atomicsOpcode(.i32_atomic_rmw_cmpxchg));
try leb.writeULEB128(writer, @as(u32, 2)); // alignment
try leb.writeULEB128(writer, @as(u32, 0)); // offset
// based on the value from the atomic check, jump to the label.
try writer.writeByte(std.wasm.opcode(.br_table));
try leb.writeULEB128(writer, @as(u32, 2)); // length of the table (we have 3 blocks but because of the mandatory default the length is 2).
try leb.writeULEB128(writer, @as(u32, 0)); // $init
try leb.writeULEB128(writer, @as(u32, 1)); // $wait
try leb.writeULEB128(writer, @as(u32, 2)); // $drop
try writer.writeByte(std.wasm.opcode(.end));
}
var it = wasm.data_segments.iterator();
var segment_index: u32 = 0;
while (it.next()) |entry| : (segment_index += 1) {
const segment: Segment = wasm.segments.items[entry.value_ptr.*];
if (segment.needsPassiveInitialization(wasm.base.options.import_memory, entry.key_ptr.*)) {
// For passive BSS segments we can simple issue a memory.fill(0).
// For non-BSS segments we do a memory.init. Both these
// instructions take as their first argument the destination
// address.
try writeI32Const(writer, segment.offset);
if (wasm.base.options.shared_memory and std.mem.eql(u8, entry.key_ptr.*, ".tdata")) {
// When we initialize the TLS segment we also set the `__tls_base`
// global. This allows the runtime to use this static copy of the
// TLS data for the first/main thread.
try writeI32Const(writer, segment.offset);
try writer.writeByte(std.wasm.opcode(.global_set));
const loc = wasm.findGlobalSymbol("__tls_base").?;
try leb.writeULEB128(writer, loc.getSymbol(wasm).index);
}
try writeI32Const(writer, 0);
try writeI32Const(writer, segment.size);
try writer.writeByte(std.wasm.opcode(.misc_prefix));
if (std.mem.eql(u8, entry.key_ptr.*, ".bss")) {
// fill bss segment with zeroes
try leb.writeULEB128(writer, std.wasm.miscOpcode(.memory_fill));
} else {
// initialize the segment
try leb.writeULEB128(writer, std.wasm.miscOpcode(.memory_init));
try leb.writeULEB128(writer, segment_index);
}
try writer.writeByte(0); // memory index immediate
}
}
if (wasm.base.options.shared_memory) {
// we set the init memory flag to value '2'
try writeI32Const(writer, flag_address);
try writeI32Const(writer, 2);
try writer.writeByte(std.wasm.opcode(.atomics_prefix));
try leb.writeULEB128(writer, std.wasm.atomicsOpcode(.i32_atomic_store));
try leb.writeULEB128(writer, @as(u32, 2)); // alignment
try leb.writeULEB128(writer, @as(u32, 0)); // offset
// notify any waiters for segment initialization completion
try writeI32Const(writer, flag_address);
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeILEB128(writer, @as(i32, -1)); // number of waiters
try writer.writeByte(std.wasm.opcode(.atomics_prefix));
try leb.writeULEB128(writer, std.wasm.atomicsOpcode(.memory_atomic_notify));
try leb.writeULEB128(writer, @as(u32, 2)); // alignment
try leb.writeULEB128(writer, @as(u32, 0)); // offset
try writer.writeByte(std.wasm.opcode(.drop));
// branch and drop segments
try writer.writeByte(std.wasm.opcode(.br));
try leb.writeULEB128(writer, @as(u32, 1));
// wait for thread to initialize memory segments
try writer.writeByte(std.wasm.opcode(.end)); // end $wait
try writeI32Const(writer, flag_address);
try writeI32Const(writer, 1); // expected flag value
try writer.writeByte(std.wasm.opcode(.i64_const));
try leb.writeILEB128(writer, @as(i64, -1)); // timeout
try writer.writeByte(std.wasm.opcode(.atomics_prefix));
try leb.writeULEB128(writer, std.wasm.atomicsOpcode(.memory_atomic_wait32));
try leb.writeULEB128(writer, @as(u32, 2)); // alignment
try leb.writeULEB128(writer, @as(u32, 0)); // offset
try writer.writeByte(std.wasm.opcode(.drop));
try writer.writeByte(std.wasm.opcode(.end)); // end $drop
}
it.reset();
segment_index = 0;
while (it.next()) |entry| : (segment_index += 1) {
const name = entry.key_ptr.*;
const segment: Segment = wasm.segments.items[entry.value_ptr.*];
if (segment.needsPassiveInitialization(wasm.base.options.import_memory, name) and
!std.mem.eql(u8, name, ".bss"))
{
// The TLS region should not be dropped since its is needed
// during the initialization of each thread (__wasm_init_tls).
if (wasm.base.options.shared_memory and std.mem.eql(u8, name, ".tdata")) {
continue;
}
try writer.writeByte(std.wasm.opcode(.misc_prefix));
try leb.writeULEB128(writer, std.wasm.miscOpcode(.data_drop));
try leb.writeULEB128(writer, segment_index);
}
}
// End of the function body
try writer.writeByte(std.wasm.opcode(.end));
try wasm.createSyntheticFunction(
"__wasm_init_memory",
std.wasm.Type{ .params = &.{}, .returns = &.{} },
&function_body,
);
}
/// Constructs a synthetic function that performs runtime relocations for
/// TLS symbols. This function is called by `__wasm_init_tls`.
fn setupTLSRelocationsFunction(wasm: *Wasm) !void {
// When we have TLS GOT entries and shared memory is enabled,
// we must perform runtime relocations or else we don't create the function.
if (!wasm.base.options.shared_memory or !wasm.requiresTLSReloc()) {
return;
}
// const loc = try wasm.createSyntheticSymbol("__wasm_apply_global_tls_relocs");
var function_body = std.ArrayList(u8).init(wasm.base.allocator);
defer function_body.deinit();
const writer = function_body.writer();
// locals (we have none)
try writer.writeByte(0);
for (wasm.got_symbols.items, 0..) |got_loc, got_index| {
const sym: *Symbol = got_loc.getSymbol(wasm);
if (!sym.isTLS()) continue; // only relocate TLS symbols
if (sym.tag == .data and sym.isDefined()) {
// get __tls_base
try writer.writeByte(std.wasm.opcode(.global_get));
try leb.writeULEB128(writer, wasm.findGlobalSymbol("__tls_base").?.getSymbol(wasm).index);
// add the virtual address of the symbol
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeULEB128(writer, sym.virtual_address);
} else if (sym.tag == .function) {
@panic("TODO: relocate GOT entry of function");
} else continue;
try writer.writeByte(std.wasm.opcode(.i32_add));
try writer.writeByte(std.wasm.opcode(.global_set));
try leb.writeULEB128(writer, wasm.imported_globals_count + @as(u32, @intCast(wasm.wasm_globals.items.len + got_index)));
}
try writer.writeByte(std.wasm.opcode(.end));
try wasm.createSyntheticFunction(
"__wasm_apply_global_tls_relocs",
std.wasm.Type{ .params = &.{}, .returns = &.{} },
&function_body,
);
}
fn validateFeatures(
wasm: *const Wasm,
to_emit: *[@typeInfo(types.Feature.Tag).Enum.fields.len]bool,
emit_features_count: *u32,
) !void {
const cpu_features = wasm.base.options.target.cpu.features;
const infer = cpu_features.isEmpty(); // when the user did not define any features, we infer them from linked objects.
const known_features_count = @typeInfo(types.Feature.Tag).Enum.fields.len;
var allowed = [_]bool{false} ** known_features_count;
var used = [_]u17{0} ** known_features_count;
var disallowed = [_]u17{0} ** known_features_count;
var required = [_]u17{0} ** known_features_count;
// when false, we fail linking. We only verify this after a loop to catch all invalid features.
var valid_feature_set = true;
// will be set to true when there's any TLS segment found in any of the object files
var has_tls = false;
// When the user has given an explicit list of features to enable,
// we extract them and insert each into the 'allowed' list.
if (!infer) {
inline for (@typeInfo(std.Target.wasm.Feature).Enum.fields) |feature_field| {
if (cpu_features.isEnabled(feature_field.value)) {
allowed[feature_field.value] = true;
emit_features_count.* += 1;
}
}
}
// extract all the used, disallowed and required features from each
// linked object file so we can test them.
for (wasm.objects.items, 0..) |object, object_index| {
for (object.features) |feature| {
const value = @as(u16, @intCast(object_index)) << 1 | @as(u1, 1);
switch (feature.prefix) {
.used => {
used[@intFromEnum(feature.tag)] = value;
},
.disallowed => {
disallowed[@intFromEnum(feature.tag)] = value;
},
.required => {
required[@intFromEnum(feature.tag)] = value;
used[@intFromEnum(feature.tag)] = value;
},
}
}
for (object.segment_info) |segment| {
if (segment.isTLS()) {
has_tls = true;
}
}
}
// when we infer the features, we allow each feature found in the 'used' set
// and insert it into the 'allowed' set. When features are not inferred,
// we validate that a used feature is allowed.
for (used, 0..) |used_set, used_index| {
const is_enabled = @as(u1, @truncate(used_set)) != 0;
if (infer) {
allowed[used_index] = is_enabled;
emit_features_count.* += @intFromBool(is_enabled);
} else if (is_enabled and !allowed[used_index]) {
log.err("feature '{}' not allowed, but used by linked object", .{@as(types.Feature.Tag, @enumFromInt(used_index))});
log.err(" defined in '{s}'", .{wasm.objects.items[used_set >> 1].name});
valid_feature_set = false;
}
}
if (!valid_feature_set) {
return error.InvalidFeatureSet;
}
if (wasm.base.options.shared_memory) {
const disallowed_feature = disallowed[@intFromEnum(types.Feature.Tag.shared_mem)];
if (@as(u1, @truncate(disallowed_feature)) != 0) {
log.err(
"shared-memory is disallowed by '{s}' because it wasn't compiled with 'atomics' and 'bulk-memory' features enabled",
.{wasm.objects.items[disallowed_feature >> 1].name},
);
valid_feature_set = false;
}
for ([_]types.Feature.Tag{ .atomics, .bulk_memory }) |feature| {
if (!allowed[@intFromEnum(feature)]) {
log.err("feature '{}' is not used but is required for shared-memory", .{feature});
}
}
}
if (has_tls) {
for ([_]types.Feature.Tag{ .atomics, .bulk_memory }) |feature| {
if (!allowed[@intFromEnum(feature)]) {
log.err("feature '{}' is not used but is required for thread-local storage", .{feature});
}
}
}
// For each linked object, validate the required and disallowed features
for (wasm.objects.items) |object| {
var object_used_features = [_]bool{false} ** known_features_count;
for (object.features) |feature| {
if (feature.prefix == .disallowed) continue; // already defined in 'disallowed' set.
// from here a feature is always used
const disallowed_feature = disallowed[@intFromEnum(feature.tag)];
if (@as(u1, @truncate(disallowed_feature)) != 0) {
log.err("feature '{}' is disallowed, but used by linked object", .{feature.tag});
log.err(" disallowed by '{s}'", .{wasm.objects.items[disallowed_feature >> 1].name});
log.err(" used in '{s}'", .{object.name});
valid_feature_set = false;
}
object_used_features[@intFromEnum(feature.tag)] = true;
}
// validate the linked object file has each required feature
for (required, 0..) |required_feature, feature_index| {
const is_required = @as(u1, @truncate(required_feature)) != 0;
if (is_required and !object_used_features[feature_index]) {
log.err("feature '{}' is required but not used in linked object", .{@as(types.Feature.Tag, @enumFromInt(feature_index))});
log.err(" required by '{s}'", .{wasm.objects.items[required_feature >> 1].name});
log.err(" missing in '{s}'", .{object.name});
valid_feature_set = false;
}
}
}
if (!valid_feature_set) {
return error.InvalidFeatureSet;
}
to_emit.* = allowed;
}
/// Creates synthetic linker-symbols, but only if they are being referenced from
/// any object file. For instance, the `__heap_base` symbol will only be created,
/// if one or multiple undefined references exist. When none exist, the symbol will
/// not be created, ensuring we don't unneccesarily emit unreferenced symbols.
fn resolveLazySymbols(wasm: *Wasm) !void {
if (wasm.string_table.getOffset("__heap_base")) |name_offset| {
if (wasm.undefs.fetchSwapRemove(name_offset)) |kv| {
const loc = try wasm.createSyntheticSymbolOffset(name_offset, .data);
try wasm.discarded.putNoClobber(wasm.base.allocator, kv.value, loc);
_ = wasm.resolved_symbols.swapRemove(loc); // we don't want to emit this symbol, only use it for relocations.
}
}
if (wasm.string_table.getOffset("__heap_end")) |name_offset| {
if (wasm.undefs.fetchSwapRemove(name_offset)) |kv| {
const loc = try wasm.createSyntheticSymbolOffset(name_offset, .data);
try wasm.discarded.putNoClobber(wasm.base.allocator, kv.value, loc);
_ = wasm.resolved_symbols.swapRemove(loc);
}
}
if (!wasm.base.options.shared_memory) {
if (wasm.string_table.getOffset("__tls_base")) |name_offset| {
if (wasm.undefs.fetchSwapRemove(name_offset)) |kv| {
const loc = try wasm.createSyntheticSymbolOffset(name_offset, .global);
try wasm.discarded.putNoClobber(wasm.base.allocator, kv.value, loc);
}
}
}
if (wasm.string_table.getOffset("__zig_errors_len")) |name_offset| {
if (wasm.undefs.fetchSwapRemove(name_offset)) |kv| {
const loc = try wasm.createSyntheticSymbolOffset(name_offset, .data);
try wasm.discarded.putNoClobber(wasm.base.allocator, kv.value, loc);
_ = wasm.resolved_symbols.swapRemove(kv.value);
}
}
}
// Tries to find a global symbol by its name. Returns null when not found,
/// and its location when it is found.
pub fn findGlobalSymbol(wasm: *Wasm, name: []const u8) ?SymbolLoc {
const offset = wasm.string_table.getOffset(name) orelse return null;
return wasm.globals.get(offset);
}
fn checkUndefinedSymbols(wasm: *const Wasm) !void {
if (wasm.base.options.output_mode == .Obj) return;
if (wasm.base.options.import_symbols) return;
var found_undefined_symbols = false;
for (wasm.undefs.values()) |undef| {
const symbol = undef.getSymbol(wasm);
if (symbol.tag == .data) {
found_undefined_symbols = true;
const file_name = if (undef.file) |file_index| name: {
break :name wasm.objects.items[file_index].name;
} else wasm.name;
const symbol_name = undef.getName(wasm);
log.err("could not resolve undefined symbol '{s}'", .{symbol_name});
log.err(" defined in '{s}'", .{file_name});
}
}
if (found_undefined_symbols) {
return error.UndefinedSymbol;
}
}
pub fn deinit(wasm: *Wasm) void {
const gpa = wasm.base.allocator;
if (wasm.llvm_object) |llvm_object| llvm_object.destroy(gpa);
for (wasm.func_types.items) |*func_type| {
func_type.deinit(gpa);
}
for (wasm.segment_info.values()) |segment_info| {
gpa.free(segment_info.name);
}
for (wasm.objects.items) |*object| {
object.deinit(gpa);
}
for (wasm.archives.items) |*archive| {
archive.deinit(gpa);
}
wasm.decls.deinit(gpa);
wasm.anon_decls.deinit(gpa);
wasm.atom_types.deinit(gpa);
wasm.symbols.deinit(gpa);
wasm.symbols_free_list.deinit(gpa);
wasm.globals.deinit(gpa);
wasm.resolved_symbols.deinit(gpa);
wasm.undefs.deinit(gpa);
wasm.discarded.deinit(gpa);
wasm.symbol_atom.deinit(gpa);
wasm.export_names.deinit(gpa);
wasm.atoms.deinit(gpa);
for (wasm.managed_atoms.items) |*managed_atom| {
managed_atom.deinit(wasm);
}
wasm.managed_atoms.deinit(gpa);
wasm.segments.deinit(gpa);
wasm.data_segments.deinit(gpa);
wasm.segment_info.deinit(gpa);
wasm.objects.deinit(gpa);
wasm.archives.deinit(gpa);
// free output sections
wasm.imports.deinit(gpa);
wasm.func_types.deinit(gpa);
wasm.functions.deinit(gpa);
wasm.wasm_globals.deinit(gpa);
wasm.function_table.deinit(gpa);
wasm.tables.deinit(gpa);
wasm.init_funcs.deinit(gpa);
wasm.exports.deinit(gpa);
wasm.string_table.deinit(gpa);
wasm.synthetic_functions.deinit(gpa);
if (wasm.dwarf) |*dwarf| {
dwarf.deinit();
}
}
/// Allocates a new symbol and returns its index.
/// Will re-use slots when a symbol was freed at an earlier stage.
pub fn allocateSymbol(wasm: *Wasm) !u32 {
try wasm.symbols.ensureUnusedCapacity(wasm.base.allocator, 1);
var symbol: Symbol = .{
.name = undefined, // will be set after updateDecl
.flags = @intFromEnum(Symbol.Flag.WASM_SYM_BINDING_LOCAL),
.tag = undefined, // will be set after updateDecl
.index = undefined, // will be set after updateDecl
.virtual_address = undefined, // will be set during atom allocation
};
if (wasm.symbols_free_list.popOrNull()) |index| {
wasm.symbols.items[index] = symbol;
return index;
}
const index = @as(u32, @intCast(wasm.symbols.items.len));
wasm.symbols.appendAssumeCapacity(symbol);
return index;
}
pub fn updateFunc(wasm: *Wasm, mod: *Module, func_index: InternPool.Index, air: Air, liveness: Liveness) !void {
if (build_options.skip_non_native and builtin.object_format != .wasm) {
@panic("Attempted to compile for object format that was disabled by build configuration");
}
if (wasm.llvm_object) |llvm_object| return llvm_object.updateFunc(mod, func_index, air, liveness);
const tracy = trace(@src());
defer tracy.end();
const func = mod.funcInfo(func_index);
const decl_index = func.owner_decl;
const decl = mod.declPtr(decl_index);
const atom_index = try wasm.getOrCreateAtomForDecl(decl_index);
const atom = wasm.getAtomPtr(atom_index);
atom.clear();
// var decl_state: ?Dwarf.DeclState = if (wasm.dwarf) |*dwarf| try dwarf.initDeclState(mod, decl_index) else null;
// defer if (decl_state) |*ds| ds.deinit();
var code_writer = std.ArrayList(u8).init(wasm.base.allocator);
defer code_writer.deinit();
// const result = try codegen.generateFunction(
// &wasm.base,
// decl.srcLoc(mod),
// func,
// air,
// liveness,
// &code_writer,
// if (decl_state) |*ds| .{ .dwarf = ds } else .none,
// );
const result = try codegen.generateFunction(
&wasm.base,
decl.srcLoc(mod),
func_index,
air,
liveness,
&code_writer,
.none,
);
const code = switch (result) {
.ok => code_writer.items,
.fail => |em| {
decl.analysis = .codegen_failure;
try mod.failed_decls.put(mod.gpa, decl_index, em);
return;
},
};
// if (wasm.dwarf) |*dwarf| {
// try dwarf.commitDeclState(
// mod,
// decl_index,
// // Actual value will be written after relocation.
// // For Wasm, this is the offset relative to the code section
// // which isn't known until flush().
// 0,
// code.len,
// &decl_state.?,
// );
// }
return wasm.finishUpdateDecl(decl_index, code);
}
// Generate code for the Decl, storing it in memory to be later written to
// the file on flush().
pub fn updateDecl(wasm: *Wasm, mod: *Module, decl_index: Module.Decl.Index) !void {
if (build_options.skip_non_native and builtin.object_format != .wasm) {
@panic("Attempted to compile for object format that was disabled by build configuration");
}
if (wasm.llvm_object) |llvm_object| return llvm_object.updateDecl(mod, decl_index);
const tracy = trace(@src());
defer tracy.end();
const decl = mod.declPtr(decl_index);
if (decl.val.getFunction(mod)) |_| {
return;
} else if (decl.val.getExternFunc(mod)) |_| {
return;
}
const atom_index = try wasm.getOrCreateAtomForDecl(decl_index);
const atom = wasm.getAtomPtr(atom_index);
atom.clear();
if (decl.isExtern(mod)) {
const variable = decl.getOwnedVariable(mod).?;
const name = mod.intern_pool.stringToSlice(decl.name);
const lib_name = mod.intern_pool.stringToSliceUnwrap(variable.lib_name);
return wasm.addOrUpdateImport(name, atom.sym_index, lib_name, null);
}
const val = if (decl.val.getVariable(mod)) |variable| variable.init.toValue() else decl.val;
var code_writer = std.ArrayList(u8).init(wasm.base.allocator);
defer code_writer.deinit();
const res = try codegen.generateSymbol(
&wasm.base,
decl.srcLoc(mod),
.{ .ty = decl.ty, .val = val },
&code_writer,
.none,
.{ .parent_atom_index = atom.sym_index },
);
const code = switch (res) {
.ok => code_writer.items,
.fail => |em| {
decl.analysis = .codegen_failure;
try mod.failed_decls.put(mod.gpa, decl_index, em);
return;
},
};
return wasm.finishUpdateDecl(decl_index, code);
}
pub fn updateDeclLineNumber(wasm: *Wasm, mod: *Module, decl_index: Module.Decl.Index) !void {
if (wasm.llvm_object) |_| return;
if (wasm.dwarf) |*dw| {
const tracy = trace(@src());
defer tracy.end();
const decl = mod.declPtr(decl_index);
const decl_name = mod.intern_pool.stringToSlice(try decl.getFullyQualifiedName(mod));
log.debug("updateDeclLineNumber {s}{*}", .{ decl_name, decl });
try dw.updateDeclLineNumber(mod, decl_index);
}
}
fn finishUpdateDecl(wasm: *Wasm, decl_index: Module.Decl.Index, code: []const u8) !void {
const mod = wasm.base.options.module.?;
const decl = mod.declPtr(decl_index);
const atom_index = wasm.decls.get(decl_index).?;
const atom = wasm.getAtomPtr(atom_index);
const symbol = &wasm.symbols.items[atom.sym_index];
const full_name = mod.intern_pool.stringToSlice(try decl.getFullyQualifiedName(mod));
symbol.name = try wasm.string_table.put(wasm.base.allocator, full_name);
try atom.code.appendSlice(wasm.base.allocator, code);
try wasm.resolved_symbols.put(wasm.base.allocator, atom.symbolLoc(), {});
atom.size = @intCast(code.len);
if (code.len == 0) return;
atom.alignment = decl.getAlignment(mod);
}
/// From a given symbol location, returns its `wasm.GlobalType`.
/// Asserts the Symbol represents a global.
fn getGlobalType(wasm: *const Wasm, loc: SymbolLoc) std.wasm.GlobalType {
const symbol = loc.getSymbol(wasm);
assert(symbol.tag == .global);
const is_undefined = symbol.isUndefined();
if (loc.file) |file_index| {
const obj: Object = wasm.objects.items[file_index];
if (is_undefined) {
return obj.findImport(.global, symbol.index).kind.global;
}
const import_global_count = obj.importedCountByKind(.global);
return obj.globals[symbol.index - import_global_count].global_type;
}
if (is_undefined) {
return wasm.imports.get(loc).?.kind.global;
}
return wasm.wasm_globals.items[symbol.index].global_type;
}
/// From a given symbol location, returns its `wasm.Type`.
/// Asserts the Symbol represents a function.
fn getFunctionSignature(wasm: *const Wasm, loc: SymbolLoc) std.wasm.Type {
const symbol = loc.getSymbol(wasm);
assert(symbol.tag == .function);
const is_undefined = symbol.isUndefined();
if (loc.file) |file_index| {
const obj: Object = wasm.objects.items[file_index];
if (is_undefined) {
const ty_index = obj.findImport(.function, symbol.index).kind.function;
return obj.func_types[ty_index];
}
const import_function_count = obj.importedCountByKind(.function);
const type_index = obj.functions[symbol.index - import_function_count].type_index;
return obj.func_types[type_index];
}
if (is_undefined) {
const ty_index = wasm.imports.get(loc).?.kind.function;
return wasm.func_types.items[ty_index];
}
return wasm.func_types.items[wasm.functions.get(.{ .file = loc.file, .index = loc.index }).?.type_index];
}
/// Lowers a constant typed value to a local symbol and atom.
/// Returns the symbol index of the local
/// The given `decl` is the parent decl whom owns the constant.
pub fn lowerUnnamedConst(wasm: *Wasm, tv: TypedValue, decl_index: Module.Decl.Index) !u32 {
const mod = wasm.base.options.module.?;
assert(tv.ty.zigTypeTag(mod) != .Fn); // cannot create local symbols for functions
const decl = mod.declPtr(decl_index);
const parent_atom_index = try wasm.getOrCreateAtomForDecl(decl_index);
const parent_atom = wasm.getAtom(parent_atom_index);
const local_index = parent_atom.locals.items.len;
const fqn = mod.intern_pool.stringToSlice(try decl.getFullyQualifiedName(mod));
const name = try std.fmt.allocPrintZ(wasm.base.allocator, "__unnamed_{s}_{d}", .{
fqn, local_index,
});
defer wasm.base.allocator.free(name);
switch (try wasm.lowerConst(name, tv, decl.srcLoc(mod))) {
.ok => |atom_index| {
try wasm.getAtomPtr(parent_atom_index).locals.append(wasm.base.allocator, atom_index);
return wasm.getAtom(atom_index).getSymbolIndex().?;
},
.fail => |em| {
decl.analysis = .codegen_failure;
try mod.failed_decls.put(mod.gpa, decl_index, em);
return error.CodegenFail;
},
}
}
const LowerConstResult = union(enum) {
ok: Atom.Index,
fail: *Module.ErrorMsg,
};
fn lowerConst(wasm: *Wasm, name: []const u8, tv: TypedValue, src_loc: Module.SrcLoc) !LowerConstResult {
const mod = wasm.base.options.module.?;
// Create and initialize a new local symbol and atom
const atom_index = try wasm.createAtom();
var value_bytes = std.ArrayList(u8).init(wasm.base.allocator);
defer value_bytes.deinit();
const code = code: {
const atom = wasm.getAtomPtr(atom_index);
atom.alignment = tv.ty.abiAlignment(mod);
wasm.symbols.items[atom.sym_index] = .{
.name = try wasm.string_table.put(wasm.base.allocator, name),
.flags = @intFromEnum(Symbol.Flag.WASM_SYM_BINDING_LOCAL),
.tag = .data,
.index = undefined,
.virtual_address = undefined,
};
try wasm.resolved_symbols.putNoClobber(wasm.base.allocator, atom.symbolLoc(), {});
const result = try codegen.generateSymbol(
&wasm.base,
src_loc,
tv,
&value_bytes,
.none,
.{
.parent_atom_index = atom.sym_index,
.addend = null,
},
);
break :code switch (result) {
.ok => value_bytes.items,
.fail => |em| {
return .{ .fail = em };
},
};
};
const atom = wasm.getAtomPtr(atom_index);
atom.size = @intCast(code.len);
try atom.code.appendSlice(wasm.base.allocator, code);
return .{ .ok = atom_index };
}
/// Returns the symbol index from a symbol of which its flag is set global,
/// such as an exported or imported symbol.
/// If the symbol does not yet exist, creates a new one symbol instead
/// and then returns the index to it.
pub fn getGlobalSymbol(wasm: *Wasm, name: []const u8, lib_name: ?[]const u8) !u32 {
_ = lib_name;
const name_index = try wasm.string_table.put(wasm.base.allocator, name);
const gop = try wasm.globals.getOrPut(wasm.base.allocator, name_index);
if (gop.found_existing) {
return gop.value_ptr.*.index;
}
var symbol: Symbol = .{
.name = name_index,
.flags = 0,
.index = undefined, // index to type will be set after merging function symbols
.tag = .function,
.virtual_address = undefined,
};
symbol.setGlobal(true);
symbol.setUndefined(true);
const sym_index = if (wasm.symbols_free_list.popOrNull()) |index| index else blk: {
var index = @as(u32, @intCast(wasm.symbols.items.len));
try wasm.symbols.ensureUnusedCapacity(wasm.base.allocator, 1);
wasm.symbols.items.len += 1;
break :blk index;
};
wasm.symbols.items[sym_index] = symbol;
gop.value_ptr.* = .{ .index = sym_index, .file = null };
try wasm.resolved_symbols.put(wasm.base.allocator, gop.value_ptr.*, {});
try wasm.undefs.putNoClobber(wasm.base.allocator, name_index, gop.value_ptr.*);
return sym_index;
}
/// For a given decl, find the given symbol index's atom, and create a relocation for the type.
/// Returns the given pointer address
pub fn getDeclVAddr(
wasm: *Wasm,
decl_index: Module.Decl.Index,
reloc_info: link.File.RelocInfo,
) !u64 {
const mod = wasm.base.options.module.?;
const decl = mod.declPtr(decl_index);
const target_atom_index = try wasm.getOrCreateAtomForDecl(decl_index);
const target_symbol_index = wasm.getAtom(target_atom_index).sym_index;
assert(reloc_info.parent_atom_index != 0);
const atom_index = wasm.symbol_atom.get(.{ .file = null, .index = reloc_info.parent_atom_index }).?;
const atom = wasm.getAtomPtr(atom_index);
const is_wasm32 = wasm.base.options.target.cpu.arch == .wasm32;
if (decl.ty.zigTypeTag(mod) == .Fn) {
assert(reloc_info.addend == 0); // addend not allowed for function relocations
// We found a function pointer, so add it to our table,
// as function pointers are not allowed to be stored inside the data section.
// They are instead stored in a function table which are called by index.
try wasm.addTableFunction(target_symbol_index);
try atom.relocs.append(wasm.base.allocator, .{
.index = target_symbol_index,
.offset = @as(u32, @intCast(reloc_info.offset)),
.relocation_type = if (is_wasm32) .R_WASM_TABLE_INDEX_I32 else .R_WASM_TABLE_INDEX_I64,
});
} else {
try atom.relocs.append(wasm.base.allocator, .{
.index = target_symbol_index,
.offset = @as(u32, @intCast(reloc_info.offset)),
.relocation_type = if (is_wasm32) .R_WASM_MEMORY_ADDR_I32 else .R_WASM_MEMORY_ADDR_I64,
.addend = @as(i32, @intCast(reloc_info.addend)),
});
}
// we do not know the final address at this point,
// as atom allocation will determine the address and relocations
// will calculate and rewrite this. Therefore, we simply return the symbol index
// that was targeted.
return target_symbol_index;
}
pub fn lowerAnonDecl(
wasm: *Wasm,
decl_val: InternPool.Index,
explicit_alignment: Alignment,
src_loc: Module.SrcLoc,
) !codegen.Result {
const gop = try wasm.anon_decls.getOrPut(wasm.base.allocator, decl_val);
if (!gop.found_existing) {
const mod = wasm.base.options.module.?;
const ty = mod.intern_pool.typeOf(decl_val).toType();
const tv: TypedValue = .{ .ty = ty, .val = decl_val.toValue() };
var name_buf: [32]u8 = undefined;
const name = std.fmt.bufPrint(&name_buf, "__anon_{d}", .{
@intFromEnum(decl_val),
}) catch unreachable;
switch (try wasm.lowerConst(name, tv, src_loc)) {
.ok => |atom_index| wasm.anon_decls.values()[gop.index] = atom_index,
.fail => |em| return .{ .fail = em },
}
}
const atom = wasm.getAtomPtr(wasm.anon_decls.values()[gop.index]);
atom.alignment = switch (atom.alignment) {
.none => explicit_alignment,
else => switch (explicit_alignment) {
.none => atom.alignment,
else => atom.alignment.maxStrict(explicit_alignment),
},
};
return .ok;
}
pub fn getAnonDeclVAddr(wasm: *Wasm, decl_val: InternPool.Index, reloc_info: link.File.RelocInfo) !u64 {
const atom_index = wasm.anon_decls.get(decl_val).?;
const target_symbol_index = wasm.getAtom(atom_index).getSymbolIndex().?;
const parent_atom_index = wasm.symbol_atom.get(.{ .file = null, .index = reloc_info.parent_atom_index }).?;
const parent_atom = wasm.getAtomPtr(parent_atom_index);
const is_wasm32 = wasm.base.options.target.cpu.arch == .wasm32;
const mod = wasm.base.options.module.?;
const ty = mod.intern_pool.typeOf(decl_val).toType();
if (ty.zigTypeTag(mod) == .Fn) {
assert(reloc_info.addend == 0); // addend not allowed for function relocations
// We found a function pointer, so add it to our table,
// as function pointers are not allowed to be stored inside the data section.
// They are instead stored in a function table which are called by index.
try wasm.addTableFunction(target_symbol_index);
try parent_atom.relocs.append(wasm.base.allocator, .{
.index = target_symbol_index,
.offset = @as(u32, @intCast(reloc_info.offset)),
.relocation_type = if (is_wasm32) .R_WASM_TABLE_INDEX_I32 else .R_WASM_TABLE_INDEX_I64,
});
} else {
try parent_atom.relocs.append(wasm.base.allocator, .{
.index = target_symbol_index,
.offset = @as(u32, @intCast(reloc_info.offset)),
.relocation_type = if (is_wasm32) .R_WASM_MEMORY_ADDR_I32 else .R_WASM_MEMORY_ADDR_I64,
.addend = @as(i32, @intCast(reloc_info.addend)),
});
}
// we do not know the final address at this point,
// as atom allocation will determine the address and relocations
// will calculate and rewrite this. Therefore, we simply return the symbol index
// that was targeted.
return target_symbol_index;
}
pub fn deleteDeclExport(wasm: *Wasm, decl_index: Module.Decl.Index) void {
if (wasm.llvm_object) |_| return;
const atom_index = wasm.decls.get(decl_index) orelse return;
const sym_index = wasm.getAtom(atom_index).sym_index;
const loc: SymbolLoc = .{ .file = null, .index = sym_index };
const symbol = loc.getSymbol(wasm);
const symbol_name = wasm.string_table.get(symbol.name);
log.debug("Deleting export for decl '{s}'", .{symbol_name});
if (wasm.export_names.fetchRemove(loc)) |kv| {
assert(wasm.globals.remove(kv.value));
} else {
assert(wasm.globals.remove(symbol.name));
}
}
pub fn updateExports(
wasm: *Wasm,
mod: *Module,
exported: Module.Exported,
exports: []const *Module.Export,
) !void {
if (build_options.skip_non_native and builtin.object_format != .wasm) {
@panic("Attempted to compile for object format that was disabled by build configuration");
}
if (wasm.llvm_object) |llvm_object| return llvm_object.updateExports(mod, exported, exports);
if (wasm.base.options.emit == null) return;
const decl_index = switch (exported) {
.decl_index => |i| i,
.value => |val| {
_ = val;
@panic("TODO: implement Wasm linker code for exporting a constant value");
},
};
const decl = mod.declPtr(decl_index);
const atom_index = try wasm.getOrCreateAtomForDecl(decl_index);
const atom = wasm.getAtom(atom_index);
const atom_sym = atom.symbolLoc().getSymbol(wasm).*;
const gpa = mod.gpa;
for (exports) |exp| {
if (mod.intern_pool.stringToSliceUnwrap(exp.opts.section)) |section| {
try mod.failed_exports.putNoClobber(gpa, exp, try Module.ErrorMsg.create(
gpa,
decl.srcLoc(mod),
"Unimplemented: ExportOptions.section '{s}'",
.{section},
));
continue;
}
const exported_decl_index = switch (exp.exported) {
.value => {
try mod.failed_exports.putNoClobber(gpa, exp, try Module.ErrorMsg.create(
gpa,
decl.srcLoc(mod),
"Unimplemented: exporting a named constant value",
.{},
));
continue;
},
.decl_index => |i| i,
};
const exported_atom_index = try wasm.getOrCreateAtomForDecl(exported_decl_index);
const exported_atom = wasm.getAtom(exported_atom_index);
const export_name = try wasm.string_table.put(wasm.base.allocator, mod.intern_pool.stringToSlice(exp.opts.name));
const sym_loc = exported_atom.symbolLoc();
const symbol = sym_loc.getSymbol(wasm);
symbol.setGlobal(true);
symbol.setUndefined(false);
symbol.index = atom_sym.index;
symbol.tag = atom_sym.tag;
symbol.name = atom_sym.name;
switch (exp.opts.linkage) {
.Internal => {
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
symbol.setFlag(.WASM_SYM_BINDING_WEAK);
},
.Weak => {
symbol.setFlag(.WASM_SYM_BINDING_WEAK);
},
.Strong => {}, // symbols are strong by default
.LinkOnce => {
try mod.failed_exports.putNoClobber(gpa, exp, try Module.ErrorMsg.create(
gpa,
decl.srcLoc(mod),
"Unimplemented: LinkOnce",
.{},
));
continue;
},
}
if (wasm.globals.get(export_name)) |existing_loc| {
if (existing_loc.index == atom.sym_index) continue;
const existing_sym: Symbol = existing_loc.getSymbol(wasm).*;
if (!existing_sym.isUndefined()) blk: {
if (symbol.isWeak()) {
try wasm.discarded.put(wasm.base.allocator, existing_loc, sym_loc);
continue; // to-be-exported symbol is weak, so we keep the existing symbol
}
// new symbol is not weak while existing is, replace existing symbol
if (existing_sym.isWeak()) {
break :blk;
}
// When both the to-be-exported symbol and the already existing symbol
// are strong symbols, we have a linker error.
// In the other case we replace one with the other.
try mod.failed_exports.put(gpa, exp, try Module.ErrorMsg.create(
gpa,
decl.srcLoc(mod),
\\LinkError: symbol '{}' defined multiple times
\\ first definition in '{s}'
\\ next definition in '{s}'
,
.{ exp.opts.name.fmt(&mod.intern_pool), wasm.name, wasm.name },
));
continue;
}
// in this case the existing symbol must be replaced either because it's weak or undefined.
try wasm.discarded.put(wasm.base.allocator, existing_loc, sym_loc);
_ = wasm.imports.remove(existing_loc);
_ = wasm.undefs.swapRemove(existing_sym.name);
}
// Ensure the symbol will be exported using the given name
if (!mod.intern_pool.stringEqlSlice(exp.opts.name, sym_loc.getName(wasm))) {
try wasm.export_names.put(wasm.base.allocator, sym_loc, export_name);
}
try wasm.globals.put(
wasm.base.allocator,
export_name,
sym_loc,
);
}
}
pub fn freeDecl(wasm: *Wasm, decl_index: Module.Decl.Index) void {
if (wasm.llvm_object) |llvm_object| return llvm_object.freeDecl(decl_index);
const mod = wasm.base.options.module.?;
const decl = mod.declPtr(decl_index);
const atom_index = wasm.decls.get(decl_index).?;
const atom = wasm.getAtomPtr(atom_index);
wasm.symbols_free_list.append(wasm.base.allocator, atom.sym_index) catch {};
_ = wasm.decls.remove(decl_index);
wasm.symbols.items[atom.sym_index].tag = .dead;
for (atom.locals.items) |local_atom_index| {
const local_atom = wasm.getAtom(local_atom_index);
const local_symbol = &wasm.symbols.items[local_atom.sym_index];
local_symbol.tag = .dead; // also for any local symbol
wasm.symbols_free_list.append(wasm.base.allocator, local_atom.sym_index) catch {};
assert(wasm.resolved_symbols.swapRemove(local_atom.symbolLoc()));
assert(wasm.symbol_atom.remove(local_atom.symbolLoc()));
}
if (decl.isExtern(mod)) {
_ = wasm.imports.remove(atom.symbolLoc());
}
_ = wasm.resolved_symbols.swapRemove(atom.symbolLoc());
_ = wasm.symbol_atom.remove(atom.symbolLoc());
// if (wasm.dwarf) |*dwarf| {
// dwarf.freeDecl(decl_index);
// }
if (atom.next) |next_atom_index| {
const next_atom = wasm.getAtomPtr(next_atom_index);
next_atom.prev = atom.prev;
atom.next = null;
}
if (atom.prev) |prev_index| {
const prev_atom = wasm.getAtomPtr(prev_index);
prev_atom.next = atom.next;
atom.prev = null;
}
}
/// Appends a new entry to the indirect function table
pub fn addTableFunction(wasm: *Wasm, symbol_index: u32) !void {
const index = @as(u32, @intCast(wasm.function_table.count()));
try wasm.function_table.put(wasm.base.allocator, .{ .file = null, .index = symbol_index }, index);
}
/// Assigns indexes to all indirect functions.
/// Starts at offset 1, where the value `0` represents an unresolved function pointer
/// or null-pointer
fn mapFunctionTable(wasm: *Wasm) void {
var it = wasm.function_table.valueIterator();
var index: u32 = 1;
while (it.next()) |value_ptr| : (index += 1) {
value_ptr.* = index;
}
if (wasm.base.options.import_table or wasm.base.options.output_mode == .Obj) {
const sym_loc = wasm.findGlobalSymbol("__indirect_function_table").?;
const import = wasm.imports.getPtr(sym_loc).?;
import.kind.table.limits.min = index - 1; // we start at index 1.
} else if (index > 1) {
log.debug("Appending indirect function table", .{});
const sym_loc = wasm.findGlobalSymbol("__indirect_function_table").?;
const symbol = sym_loc.getSymbol(wasm);
const table = &wasm.tables.items[symbol.index - wasm.imported_tables_count];
table.limits = .{ .min = index, .max = index, .flags = 0x1 };
}
}
/// Either creates a new import, or updates one if existing.
/// When `type_index` is non-null, we assume an external function.
/// In all other cases, a data-symbol will be created instead.
pub fn addOrUpdateImport(
wasm: *Wasm,
/// Name of the import
name: []const u8,
/// Symbol index that is external
symbol_index: u32,
/// Optional library name (i.e. `extern "c" fn foo() void`
lib_name: ?[:0]const u8,
/// The index of the type that represents the function signature
/// when the extern is a function. When this is null, a data-symbol
/// is asserted instead.
type_index: ?u32,
) !void {
assert(symbol_index != 0);
// For the import name, we use the decl's name, rather than the fully qualified name
// Also mangle the name when the lib name is set and not equal to "C" so imports with the same
// name but different module can be resolved correctly.
const mangle_name = lib_name != null and
!std.mem.eql(u8, lib_name.?, "c");
const full_name = if (mangle_name) full_name: {
break :full_name try std.fmt.allocPrint(wasm.base.allocator, "{s}|{s}", .{ name, lib_name.? });
} else name;
defer if (mangle_name) wasm.base.allocator.free(full_name);
const decl_name_index = try wasm.string_table.put(wasm.base.allocator, full_name);
const symbol: *Symbol = &wasm.symbols.items[symbol_index];
symbol.setUndefined(true);
symbol.setGlobal(true);
symbol.name = decl_name_index;
if (mangle_name) {
// we specified a specific name for the symbol that does not match the import name
symbol.setFlag(.WASM_SYM_EXPLICIT_NAME);
}
const global_gop = try wasm.globals.getOrPut(wasm.base.allocator, decl_name_index);
if (!global_gop.found_existing) {
const loc: SymbolLoc = .{ .file = null, .index = symbol_index };
global_gop.value_ptr.* = loc;
try wasm.resolved_symbols.put(wasm.base.allocator, loc, {});
try wasm.undefs.putNoClobber(wasm.base.allocator, decl_name_index, loc);
} else if (global_gop.value_ptr.*.index != symbol_index) {
// We are not updating a symbol, but found an existing global
// symbol with the same name. This means we always favor the
// existing symbol, regardless whether it's defined or not.
// We can also skip storing the import as we will not output
// this symbol.
return wasm.discarded.put(
wasm.base.allocator,
.{ .file = null, .index = symbol_index },
global_gop.value_ptr.*,
);
}
if (type_index) |ty_index| {
const gop = try wasm.imports.getOrPut(wasm.base.allocator, .{ .index = symbol_index, .file = null });
const module_name = if (lib_name) |l_name| blk: {
break :blk l_name;
} else wasm.host_name;
if (!gop.found_existing) {
gop.value_ptr.* = .{
.module_name = try wasm.string_table.put(wasm.base.allocator, module_name),
.name = try wasm.string_table.put(wasm.base.allocator, name),
.kind = .{ .function = ty_index },
};
}
} else {
// non-functions will not be imported from the runtime, but only resolved during link-time
symbol.tag = .data;
}
}
/// Kind represents the type of an Atom, which is only
/// used to parse a decl into an Atom to define in which section
/// or segment it should be placed.
const Kind = union(enum) {
/// Represents the segment the data symbol should
/// be inserted into.
/// TODO: Add TLS segments
data: enum {
read_only,
uninitialized,
initialized,
},
function: void,
/// Returns the segment name the data kind represents.
/// Asserts `kind` has its active tag set to `data`.
fn segmentName(kind: Kind) []const u8 {
switch (kind.data) {
.read_only => return ".rodata.",
.uninitialized => return ".bss.",
.initialized => return ".data.",
}
}
};
/// Parses an Atom and inserts its metadata into the corresponding sections.
fn parseAtom(wasm: *Wasm, atom_index: Atom.Index, kind: Kind) !void {
const atom = wasm.getAtomPtr(atom_index);
const symbol = (SymbolLoc{ .file = null, .index = atom.sym_index }).getSymbol(wasm);
const final_index: u32 = switch (kind) {
.function => result: {
const index = @as(u32, @intCast(wasm.functions.count() + wasm.imported_functions_count));
const type_index = wasm.atom_types.get(atom_index).?;
try wasm.functions.putNoClobber(
wasm.base.allocator,
.{ .file = null, .index = index },
.{ .type_index = type_index },
);
symbol.tag = .function;
symbol.index = index;
if (wasm.code_section_index == null) {
wasm.code_section_index = @as(u32, @intCast(wasm.segments.items.len));
try wasm.segments.append(wasm.base.allocator, .{
.alignment = atom.alignment,
.size = atom.size,
.offset = 0,
.flags = 0,
});
}
break :result wasm.code_section_index.?;
},
.data => result: {
const segment_name = try std.mem.concat(wasm.base.allocator, u8, &.{
kind.segmentName(),
wasm.string_table.get(symbol.name),
});
errdefer wasm.base.allocator.free(segment_name);
const segment_info: types.Segment = .{
.name = segment_name,
.alignment = atom.alignment,
.flags = 0,
};
symbol.tag = .data;
// when creating an object file, or importing memory and the data belongs in the .bss segment
// we set the entire region of it to zeroes.
// We do not have to do this when exporting the memory (the default) because the runtime
// will do it for us, and we do not emit the bss segment at all.
if ((wasm.base.options.output_mode == .Obj or wasm.base.options.import_memory) and kind.data == .uninitialized) {
@memset(atom.code.items, 0);
}
const should_merge = wasm.base.options.output_mode != .Obj;
const gop = try wasm.data_segments.getOrPut(wasm.base.allocator, segment_info.outputName(should_merge));
if (gop.found_existing) {
const index = gop.value_ptr.*;
wasm.segments.items[index].size += atom.size;
symbol.index = @as(u32, @intCast(wasm.segment_info.getIndex(index).?));
// segment info already exists, so free its memory
wasm.base.allocator.free(segment_name);
break :result index;
} else {
const index = @as(u32, @intCast(wasm.segments.items.len));
var flags: u32 = 0;
if (wasm.base.options.shared_memory) {
flags |= @intFromEnum(Segment.Flag.WASM_DATA_SEGMENT_IS_PASSIVE);
}
try wasm.segments.append(wasm.base.allocator, .{
.alignment = atom.alignment,
.size = 0,
.offset = 0,
.flags = flags,
});
gop.value_ptr.* = index;
const info_index = @as(u32, @intCast(wasm.segment_info.count()));
try wasm.segment_info.put(wasm.base.allocator, index, segment_info);
symbol.index = info_index;
break :result index;
}
},
};
const segment: *Segment = &wasm.segments.items[final_index];
segment.alignment = segment.alignment.max(atom.alignment);
try wasm.appendAtomAtIndex(final_index, atom_index);
}
/// From a given index, append the given `Atom` at the back of the linked list.
/// Simply inserts it into the map of atoms when it doesn't exist yet.
pub fn appendAtomAtIndex(wasm: *Wasm, index: u32, atom_index: Atom.Index) !void {
const atom = wasm.getAtomPtr(atom_index);
if (wasm.atoms.getPtr(index)) |last_index_ptr| {
const last = wasm.getAtomPtr(last_index_ptr.*);
last.*.next = atom_index;
atom.prev = last_index_ptr.*;
last_index_ptr.* = atom_index;
} else {
try wasm.atoms.putNoClobber(wasm.base.allocator, index, atom_index);
}
}
/// Allocates debug atoms into their respective debug sections
/// to merge them with maybe-existing debug atoms from object files.
fn allocateDebugAtoms(wasm: *Wasm) !void {
if (wasm.dwarf == null) return;
const allocAtom = struct {
fn f(bin: *Wasm, maybe_index: *?u32, atom_index: Atom.Index) !void {
const index = maybe_index.* orelse idx: {
const index = @as(u32, @intCast(bin.segments.items.len));
try bin.appendDummySegment();
maybe_index.* = index;
break :idx index;
};
const atom = bin.getAtomPtr(atom_index);
atom.size = @as(u32, @intCast(atom.code.items.len));
bin.symbols.items[atom.sym_index].index = index;
try bin.appendAtomAtIndex(index, atom_index);
}
}.f;
try allocAtom(wasm, &wasm.debug_info_index, wasm.debug_info_atom.?);
try allocAtom(wasm, &wasm.debug_line_index, wasm.debug_line_atom.?);
try allocAtom(wasm, &wasm.debug_loc_index, wasm.debug_loc_atom.?);
try allocAtom(wasm, &wasm.debug_str_index, wasm.debug_str_atom.?);
try allocAtom(wasm, &wasm.debug_ranges_index, wasm.debug_ranges_atom.?);
try allocAtom(wasm, &wasm.debug_abbrev_index, wasm.debug_abbrev_atom.?);
try allocAtom(wasm, &wasm.debug_pubnames_index, wasm.debug_pubnames_atom.?);
try allocAtom(wasm, &wasm.debug_pubtypes_index, wasm.debug_pubtypes_atom.?);
}
fn allocateAtoms(wasm: *Wasm) !void {
// first sort the data segments
try sortDataSegments(wasm);
try allocateDebugAtoms(wasm);
var it = wasm.atoms.iterator();
while (it.next()) |entry| {
const segment = &wasm.segments.items[entry.key_ptr.*];
var atom_index = entry.value_ptr.*;
var offset: u32 = 0;
while (true) {
const atom = wasm.getAtomPtr(atom_index);
const symbol_loc = atom.symbolLoc();
if (wasm.code_section_index) |index| {
if (index == entry.key_ptr.*) {
if (!wasm.resolved_symbols.contains(symbol_loc)) {
// only allocate resolved function body's.
atom_index = atom.prev orelse break;
continue;
}
}
}
offset = @intCast(atom.alignment.forward(offset));
atom.offset = offset;
log.debug("Atom '{s}' allocated from 0x{x:0>8} to 0x{x:0>8} size={d}", .{
symbol_loc.getName(wasm),
offset,
offset + atom.size,
atom.size,
});
offset += atom.size;
atom_index = atom.prev orelse break;
}
segment.size = @intCast(segment.alignment.forward(offset));
}
}
/// For each data symbol, sets the virtual address.
fn allocateVirtualAddresses(wasm: *Wasm) void {
for (wasm.resolved_symbols.keys()) |loc| {
const symbol = loc.getSymbol(wasm);
if (symbol.tag != .data) {
continue; // only data symbols have virtual addresses
}
const atom_index = wasm.symbol_atom.get(loc) orelse {
// synthetic symbol that does not contain an atom
continue;
};
const atom = wasm.getAtom(atom_index);
const merge_segment = wasm.base.options.output_mode != .Obj;
const segment_info = if (atom.file) |object_index| blk: {
break :blk wasm.objects.items[object_index].segment_info;
} else wasm.segment_info.values();
const segment_name = segment_info[symbol.index].outputName(merge_segment);
const segment_index = wasm.data_segments.get(segment_name).?;
const segment = wasm.segments.items[segment_index];
// TLS symbols have their virtual address set relative to their own TLS segment,
// rather than the entire Data section.
if (symbol.hasFlag(.WASM_SYM_TLS)) {
symbol.virtual_address = atom.offset;
} else {
symbol.virtual_address = atom.offset + segment.offset;
}
}
}
fn sortDataSegments(wasm: *Wasm) !void {
var new_mapping: std.StringArrayHashMapUnmanaged(u32) = .{};
try new_mapping.ensureUnusedCapacity(wasm.base.allocator, wasm.data_segments.count());
errdefer new_mapping.deinit(wasm.base.allocator);
const keys = try wasm.base.allocator.dupe([]const u8, wasm.data_segments.keys());
defer wasm.base.allocator.free(keys);
const SortContext = struct {
fn sort(_: void, lhs: []const u8, rhs: []const u8) bool {
return order(lhs) < order(rhs);
}
fn order(name: []const u8) u8 {
if (mem.startsWith(u8, name, ".rodata")) return 0;
if (mem.startsWith(u8, name, ".data")) return 1;
if (mem.startsWith(u8, name, ".text")) return 2;
return 3;
}
};
mem.sort([]const u8, keys, {}, SortContext.sort);
for (keys) |key| {
const segment_index = wasm.data_segments.get(key).?;
new_mapping.putAssumeCapacity(key, segment_index);
}
wasm.data_segments.deinit(wasm.base.allocator);
wasm.data_segments = new_mapping;
}
/// Obtains all initfuncs from each object file, verifies its function signature,
/// and then appends it to our final `init_funcs` list.
/// After all functions have been inserted, the functions will be ordered based
/// on their priority.
/// NOTE: This function must be called before we merged any other section.
/// This is because all init funcs in the object files contain references to the
/// original functions and their types. We need to know the type to verify it doesn't
/// contain any parameters.
fn setupInitFunctions(wasm: *Wasm) !void {
for (wasm.objects.items, 0..) |object, file_index| {
try wasm.init_funcs.ensureUnusedCapacity(wasm.base.allocator, object.init_funcs.len);
for (object.init_funcs) |init_func| {
const symbol = object.symtable[init_func.symbol_index];
const ty: std.wasm.Type = if (symbol.isUndefined()) ty: {
const imp: types.Import = object.findImport(.function, symbol.index);
break :ty object.func_types[imp.kind.function];
} else ty: {
const func_index = symbol.index - object.importedCountByKind(.function);
const func = object.functions[func_index];
break :ty object.func_types[func.type_index];
};
if (ty.params.len != 0) {
log.err("constructor functions cannot take arguments: '{s}'", .{object.string_table.get(symbol.name)});
return error.InvalidInitFunc;
}
log.debug("appended init func '{s}'\n", .{object.string_table.get(symbol.name)});
wasm.init_funcs.appendAssumeCapacity(.{
.index = init_func.symbol_index,
.file = @as(u16, @intCast(file_index)),
.priority = init_func.priority,
});
}
}
// sort the initfunctions based on their priority
mem.sort(InitFuncLoc, wasm.init_funcs.items, {}, InitFuncLoc.lessThan);
}
/// Generates an atom containing the global error set' size.
/// This will only be generated if the symbol exists.
fn setupErrorsLen(wasm: *Wasm) !void {
const loc = wasm.findGlobalSymbol("__zig_errors_len") orelse return;
const errors_len = wasm.base.options.module.?.global_error_set.count();
// overwrite existing atom if it already exists (maybe the error set has increased)
// if not, allcoate a new atom.
const atom_index = if (wasm.symbol_atom.get(loc)) |index| blk: {
const atom = wasm.getAtomPtr(index);
if (atom.next) |next_atom_index| {
const next_atom = wasm.getAtomPtr(next_atom_index);
next_atom.prev = atom.prev;
atom.next = null;
}
if (atom.prev) |prev_index| {
const prev_atom = wasm.getAtomPtr(prev_index);
prev_atom.next = atom.next;
atom.prev = null;
}
atom.deinit(wasm);
break :blk index;
} else new_atom: {
const atom_index: Atom.Index = @intCast(wasm.managed_atoms.items.len);
try wasm.symbol_atom.put(wasm.base.allocator, loc, atom_index);
try wasm.managed_atoms.append(wasm.base.allocator, undefined);
break :new_atom atom_index;
};
const atom = wasm.getAtomPtr(atom_index);
atom.* = Atom.empty;
atom.sym_index = loc.index;
atom.size = 2;
try atom.code.writer(wasm.base.allocator).writeInt(u16, @intCast(errors_len), .little);
try wasm.parseAtom(atom_index, .{ .data = .read_only });
}
/// Creates a function body for the `__wasm_call_ctors` symbol.
/// Loops over all constructors found in `init_funcs` and calls them
/// respectively based on their priority which was sorted by `setupInitFunctions`.
/// NOTE: This function must be called after we merged all sections to ensure the
/// references to the function stored in the symbol have been finalized so we end
/// up calling the resolved function.
fn initializeCallCtorsFunction(wasm: *Wasm) !void {
// No code to emit, so also no ctors to call
if (wasm.code_section_index == null) {
// Make sure to remove it from the resolved symbols so we do not emit
// it within any section. TODO: Remove this once we implement garbage collection.
const loc = wasm.findGlobalSymbol("__wasm_call_ctors").?;
std.debug.assert(wasm.resolved_symbols.swapRemove(loc));
return;
}
var function_body = std.ArrayList(u8).init(wasm.base.allocator);
defer function_body.deinit();
const writer = function_body.writer();
// Create the function body
{
// Write locals count (we have none)
try leb.writeULEB128(writer, @as(u32, 0));
// call constructors
for (wasm.init_funcs.items) |init_func_loc| {
const symbol = init_func_loc.getSymbol(wasm);
const func = wasm.functions.values()[symbol.index - wasm.imported_functions_count];
const ty = wasm.func_types.items[func.type_index];
// Call function by its function index
try writer.writeByte(std.wasm.opcode(.call));
try leb.writeULEB128(writer, symbol.index);
// drop all returned values from the stack as __wasm_call_ctors has no return value
for (ty.returns) |_| {
try writer.writeByte(std.wasm.opcode(.drop));
}
}
// End function body
try writer.writeByte(std.wasm.opcode(.end));
}
try wasm.createSyntheticFunction(
"__wasm_call_ctors",
std.wasm.Type{ .params = &.{}, .returns = &.{} },
&function_body,
);
}
fn createSyntheticFunction(
wasm: *Wasm,
symbol_name: []const u8,
func_ty: std.wasm.Type,
function_body: *std.ArrayList(u8),
) !void {
const loc = wasm.findGlobalSymbol(symbol_name) orelse
try wasm.createSyntheticSymbol(symbol_name, .function);
const symbol = loc.getSymbol(wasm);
const ty_index = try wasm.putOrGetFuncType(func_ty);
// create function with above type
const func_index = wasm.imported_functions_count + @as(u32, @intCast(wasm.functions.count()));
try wasm.functions.putNoClobber(
wasm.base.allocator,
.{ .file = null, .index = func_index },
.{ .type_index = ty_index },
);
symbol.index = func_index;
// create the atom that will be output into the final binary
const atom_index = @as(Atom.Index, @intCast(wasm.managed_atoms.items.len));
const atom = try wasm.managed_atoms.addOne(wasm.base.allocator);
atom.* = .{
.size = @as(u32, @intCast(function_body.items.len)),
.offset = 0,
.sym_index = loc.index,
.file = null,
.alignment = .@"1",
.next = null,
.prev = null,
.code = function_body.moveToUnmanaged(),
};
try wasm.appendAtomAtIndex(wasm.code_section_index.?, atom_index);
try wasm.symbol_atom.putNoClobber(wasm.base.allocator, loc, atom_index);
// `allocateAtoms` has already been called, set the atom's offset manually.
// This is fine to do manually as we insert the atom at the very end.
const prev_atom = wasm.getAtom(atom.prev.?);
atom.offset = prev_atom.offset + prev_atom.size;
}
/// Unlike `createSyntheticFunction` this function is to be called by
/// the codegeneration backend. This will not allocate the created Atom yet,
/// but will instead be appended to `synthetic_functions` list and will be
/// parsed at the end of code generation.
/// Returns the index of the symbol.
pub fn createFunction(
wasm: *Wasm,
symbol_name: []const u8,
func_ty: std.wasm.Type,
function_body: *std.ArrayList(u8),
relocations: *std.ArrayList(Relocation),
) !u32 {
const loc = try wasm.createSyntheticSymbol(symbol_name, .function);
const atom_index = @as(Atom.Index, @intCast(wasm.managed_atoms.items.len));
const atom = try wasm.managed_atoms.addOne(wasm.base.allocator);
atom.* = .{
.size = @intCast(function_body.items.len),
.offset = 0,
.sym_index = loc.index,
.file = null,
.alignment = .@"1",
.next = null,
.prev = null,
.code = function_body.moveToUnmanaged(),
.relocs = relocations.moveToUnmanaged(),
};
const symbol = loc.getSymbol(wasm);
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN); // ensure function does not get exported
const section_index = wasm.code_section_index orelse idx: {
const index = @as(u32, @intCast(wasm.segments.items.len));
try wasm.appendDummySegment();
break :idx index;
};
try wasm.appendAtomAtIndex(section_index, atom_index);
try wasm.symbol_atom.putNoClobber(wasm.base.allocator, loc, atom_index);
try wasm.atom_types.put(wasm.base.allocator, atom_index, try wasm.putOrGetFuncType(func_ty));
try wasm.synthetic_functions.append(wasm.base.allocator, atom_index);
return loc.index;
}
/// If required, sets the function index in the `start` section.
fn setupStartSection(wasm: *Wasm) !void {
if (wasm.findGlobalSymbol("__wasm_init_memory")) |loc| {
wasm.entry = loc.getSymbol(wasm).index;
}
}
fn initializeTLSFunction(wasm: *Wasm) !void {
if (!wasm.base.options.shared_memory) return;
var function_body = std.ArrayList(u8).init(wasm.base.allocator);
defer function_body.deinit();
const writer = function_body.writer();
// locals
try writer.writeByte(0);
// If there's a TLS segment, initialize it during runtime using the bulk-memory feature
if (wasm.data_segments.getIndex(".tdata")) |data_index| {
const segment_index = wasm.data_segments.entries.items(.value)[data_index];
const segment = wasm.segments.items[segment_index];
const param_local: u32 = 0;
try writer.writeByte(std.wasm.opcode(.local_get));
try leb.writeULEB128(writer, param_local);
const tls_base_loc = wasm.findGlobalSymbol("__tls_base").?;
try writer.writeByte(std.wasm.opcode(.global_set));
try leb.writeULEB128(writer, tls_base_loc.getSymbol(wasm).index);
// load stack values for the bulk-memory operation
{
try writer.writeByte(std.wasm.opcode(.local_get));
try leb.writeULEB128(writer, param_local);
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeULEB128(writer, @as(u32, 0)); //segment offset
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeULEB128(writer, @as(u32, segment.size)); //segment offset
}
// perform the bulk-memory operation to initialize the data segment
try writer.writeByte(std.wasm.opcode(.misc_prefix));
try leb.writeULEB128(writer, std.wasm.miscOpcode(.memory_init));
// segment immediate
try leb.writeULEB128(writer, @as(u32, @intCast(data_index)));
// memory index immediate (always 0)
try leb.writeULEB128(writer, @as(u32, 0));
}
// If we have to perform any TLS relocations, call the corresponding function
// which performs all runtime TLS relocations. This is a synthetic function,
// generated by the linker.
if (wasm.findGlobalSymbol("__wasm_apply_global_tls_relocs")) |loc| {
try writer.writeByte(std.wasm.opcode(.call));
try leb.writeULEB128(writer, loc.getSymbol(wasm).index);
}
try writer.writeByte(std.wasm.opcode(.end));
try wasm.createSyntheticFunction(
"__wasm_init_tls",
std.wasm.Type{ .params = &.{.i32}, .returns = &.{} },
&function_body,
);
}
fn setupImports(wasm: *Wasm) !void {
log.debug("Merging imports", .{});
var discarded_it = wasm.discarded.keyIterator();
while (discarded_it.next()) |discarded| {
if (discarded.file == null) {
// remove an import if it was resolved
if (wasm.imports.remove(discarded.*)) {
log.debug("Removed symbol '{s}' as an import", .{
discarded.getName(wasm),
});
}
}
}
for (wasm.resolved_symbols.keys()) |symbol_loc| {
if (symbol_loc.file == null) {
// imports generated by Zig code are already in the `import` section
continue;
}
const symbol = symbol_loc.getSymbol(wasm);
if (std.mem.eql(u8, symbol_loc.getName(wasm), "__indirect_function_table")) {
continue;
}
if (!symbol.requiresImport()) {
continue;
}
log.debug("Symbol '{s}' will be imported from the host", .{symbol_loc.getName(wasm)});
const object = wasm.objects.items[symbol_loc.file.?];
const import = object.findImport(symbol.tag.externalType(), symbol.index);
// We copy the import to a new import to ensure the names contain references
// to the internal string table, rather than of the object file.
var new_imp: types.Import = .{
.module_name = try wasm.string_table.put(wasm.base.allocator, object.string_table.get(import.module_name)),
.name = try wasm.string_table.put(wasm.base.allocator, object.string_table.get(import.name)),
.kind = import.kind,
};
// TODO: De-duplicate imports when they contain the same names and type
try wasm.imports.putNoClobber(wasm.base.allocator, symbol_loc, new_imp);
}
// Assign all indexes of the imports to their representing symbols
var function_index: u32 = 0;
var global_index: u32 = 0;
var table_index: u32 = 0;
var it = wasm.imports.iterator();
while (it.next()) |entry| {
const symbol = entry.key_ptr.*.getSymbol(wasm);
const import: types.Import = entry.value_ptr.*;
switch (import.kind) {
.function => {
symbol.index = function_index;
function_index += 1;
},
.global => {
symbol.index = global_index;
global_index += 1;
},
.table => {
symbol.index = table_index;
table_index += 1;
},
else => unreachable,
}
}
wasm.imported_functions_count = function_index;
wasm.imported_globals_count = global_index;
wasm.imported_tables_count = table_index;
log.debug("Merged ({d}) functions, ({d}) globals, and ({d}) tables into import section", .{
function_index,
global_index,
table_index,
});
}
/// Takes the global, function and table section from each linked object file
/// and merges it into a single section for each.
fn mergeSections(wasm: *Wasm) !void {
for (wasm.resolved_symbols.keys()) |sym_loc| {
if (sym_loc.file == null) {
// Zig code-generated symbols are already within the sections and do not
// require to be merged
continue;
}
const object = &wasm.objects.items[sym_loc.file.?];
const symbol = &object.symtable[sym_loc.index];
if (symbol.isUndefined() or (symbol.tag != .function and symbol.tag != .global and symbol.tag != .table)) {
// Skip undefined symbols as they go in the `import` section
// Also skip symbols that do not need to have a section merged.
continue;
}
const offset = object.importedCountByKind(symbol.tag.externalType());
const index = symbol.index - offset;
switch (symbol.tag) {
.function => {
const gop = try wasm.functions.getOrPut(
wasm.base.allocator,
.{ .file = sym_loc.file, .index = symbol.index },
);
if (!gop.found_existing) {
gop.value_ptr.* = object.functions[index];
}
symbol.index = @as(u32, @intCast(gop.index)) + wasm.imported_functions_count;
},
.global => {
const original_global = object.globals[index];
symbol.index = @as(u32, @intCast(wasm.wasm_globals.items.len)) + wasm.imported_globals_count;
try wasm.wasm_globals.append(wasm.base.allocator, original_global);
},
.table => {
const original_table = object.tables[index];
symbol.index = @as(u32, @intCast(wasm.tables.items.len)) + wasm.imported_tables_count;
try wasm.tables.append(wasm.base.allocator, original_table);
},
else => unreachable,
}
}
log.debug("Merged ({d}) functions", .{wasm.functions.count()});
log.debug("Merged ({d}) globals", .{wasm.wasm_globals.items.len});
log.debug("Merged ({d}) tables", .{wasm.tables.items.len});
}
/// Merges function types of all object files into the final
/// 'types' section, while assigning the type index to the representing
/// section (import, export, function).
fn mergeTypes(wasm: *Wasm) !void {
// A map to track which functions have already had their
// type inserted. If we do this for the same function multiple times,
// it will be overwritten with the incorrect type.
var dirty = std.AutoHashMap(u32, void).init(wasm.base.allocator);
try dirty.ensureUnusedCapacity(@as(u32, @intCast(wasm.functions.count())));
defer dirty.deinit();
for (wasm.resolved_symbols.keys()) |sym_loc| {
if (sym_loc.file == null) {
// zig code-generated symbols are already present in final type section
continue;
}
const object = wasm.objects.items[sym_loc.file.?];
const symbol = object.symtable[sym_loc.index];
if (symbol.tag != .function) {
// Only functions have types
continue;
}
if (symbol.isUndefined()) {
log.debug("Adding type from extern function '{s}'", .{sym_loc.getName(wasm)});
const import: *types.Import = wasm.imports.getPtr(sym_loc) orelse continue;
const original_type = object.func_types[import.kind.function];
import.kind.function = try wasm.putOrGetFuncType(original_type);
} else if (!dirty.contains(symbol.index)) {
log.debug("Adding type from function '{s}'", .{sym_loc.getName(wasm)});
const func = &wasm.functions.values()[symbol.index - wasm.imported_functions_count];
func.type_index = try wasm.putOrGetFuncType(object.func_types[func.type_index]);
dirty.putAssumeCapacityNoClobber(symbol.index, {});
}
}
log.debug("Completed merging and deduplicating types. Total count: ({d})", .{wasm.func_types.items.len});
}
fn setupExports(wasm: *Wasm) !void {
if (wasm.base.options.output_mode == .Obj) return;
log.debug("Building exports from symbols", .{});
const force_exp_names = wasm.base.options.export_symbol_names;
if (force_exp_names.len > 0) {
var failed_exports = false;
for (force_exp_names) |exp_name| {
const loc = wasm.findGlobalSymbol(exp_name) orelse {
log.err("could not export '{s}', symbol not found", .{exp_name});
failed_exports = true;
continue;
};
const symbol = loc.getSymbol(wasm);
symbol.setFlag(.WASM_SYM_EXPORTED);
}
if (failed_exports) {
return error.MissingSymbol;
}
}
for (wasm.resolved_symbols.keys()) |sym_loc| {
const symbol = sym_loc.getSymbol(wasm);
if (!symbol.isExported(wasm.base.options.rdynamic)) continue;
const sym_name = sym_loc.getName(wasm);
const export_name = if (wasm.export_names.get(sym_loc)) |name| name else blk: {
if (sym_loc.file == null) break :blk symbol.name;
break :blk try wasm.string_table.put(wasm.base.allocator, sym_name);
};
const exp: types.Export = if (symbol.tag == .data) exp: {
const global_index = @as(u32, @intCast(wasm.imported_globals_count + wasm.wasm_globals.items.len));
try wasm.wasm_globals.append(wasm.base.allocator, .{
.global_type = .{ .valtype = .i32, .mutable = false },
.init = .{ .i32_const = @as(i32, @intCast(symbol.virtual_address)) },
});
break :exp .{
.name = export_name,
.kind = .global,
.index = global_index,
};
} else .{
.name = export_name,
.kind = symbol.tag.externalType(),
.index = symbol.index,
};
log.debug("Exporting symbol '{s}' as '{s}' at index: ({d})", .{
sym_name,
wasm.string_table.get(exp.name),
exp.index,
});
try wasm.exports.append(wasm.base.allocator, exp);
}
log.debug("Completed building exports. Total count: ({d})", .{wasm.exports.items.len});
}
fn setupStart(wasm: *Wasm) !void {
const entry_name = wasm.base.options.entry orelse "_start";
const symbol_loc = wasm.findGlobalSymbol(entry_name) orelse {
if (wasm.base.options.output_mode == .Exe) {
if (wasm.base.options.wasi_exec_model == .reactor) return; // Not required for reactors
} else {
return; // No entry point needed for non-executable wasm files
}
log.err("Entry symbol '{s}' missing", .{entry_name});
return error.MissingSymbol;
};
const symbol = symbol_loc.getSymbol(wasm);
if (symbol.tag != .function) {
log.err("Entry symbol '{s}' is not a function", .{entry_name});
return error.InvalidEntryKind;
}
// Ensure the symbol is exported so host environment can access it
if (wasm.base.options.output_mode != .Obj) {
symbol.setFlag(.WASM_SYM_EXPORTED);
}
}
/// Sets up the memory section of the wasm module, as well as the stack.
fn setupMemory(wasm: *Wasm) !void {
log.debug("Setting up memory layout", .{});
const page_size = std.wasm.page_size; // 64kb
// Use the user-provided stack size or else we use 1MB by default
const stack_size = wasm.base.options.stack_size_override orelse page_size * 16;
const stack_alignment: Alignment = .@"16"; // wasm's stack alignment as specified by tool-convention
const heap_alignment: Alignment = .@"16"; // wasm's heap alignment as specified by tool-convention
// Always place the stack at the start by default
// unless the user specified the global-base flag
var place_stack_first = true;
var memory_ptr: u64 = if (wasm.base.options.global_base) |base| blk: {
place_stack_first = false;
break :blk base;
} else 0;
const is_obj = wasm.base.options.output_mode == .Obj;
if (place_stack_first and !is_obj) {
memory_ptr = stack_alignment.forward(memory_ptr);
memory_ptr += stack_size;
// We always put the stack pointer global at index 0
wasm.wasm_globals.items[0].init.i32_const = @as(i32, @bitCast(@as(u32, @intCast(memory_ptr))));
}
var offset: u32 = @as(u32, @intCast(memory_ptr));
var data_seg_it = wasm.data_segments.iterator();
while (data_seg_it.next()) |entry| {
const segment = &wasm.segments.items[entry.value_ptr.*];
memory_ptr = segment.alignment.forward(memory_ptr);
// set TLS-related symbols
if (mem.eql(u8, entry.key_ptr.*, ".tdata")) {
if (wasm.findGlobalSymbol("__tls_size")) |loc| {
const sym = loc.getSymbol(wasm);
wasm.wasm_globals.items[sym.index - wasm.imported_globals_count].init.i32_const = @intCast(segment.size);
}
if (wasm.findGlobalSymbol("__tls_align")) |loc| {
const sym = loc.getSymbol(wasm);
wasm.wasm_globals.items[sym.index - wasm.imported_globals_count].init.i32_const = @intCast(segment.alignment.toByteUnitsOptional().?);
}
if (wasm.findGlobalSymbol("__tls_base")) |loc| {
const sym = loc.getSymbol(wasm);
wasm.wasm_globals.items[sym.index - wasm.imported_globals_count].init.i32_const = if (wasm.base.options.shared_memory)
@as(i32, 0)
else
@as(i32, @intCast(memory_ptr));
}
}
memory_ptr += segment.size;
segment.offset = offset;
offset += segment.size;
}
// create the memory init flag which is used by the init memory function
if (wasm.base.options.shared_memory and wasm.hasPassiveInitializationSegments()) {
// align to pointer size
memory_ptr = mem.alignForward(u64, memory_ptr, 4);
const loc = try wasm.createSyntheticSymbol("__wasm_init_memory_flag", .data);
const sym = loc.getSymbol(wasm);
sym.virtual_address = @as(u32, @intCast(memory_ptr));
memory_ptr += 4;
}
if (!place_stack_first and !is_obj) {
memory_ptr = stack_alignment.forward(memory_ptr);
memory_ptr += stack_size;
wasm.wasm_globals.items[0].init.i32_const = @as(i32, @bitCast(@as(u32, @intCast(memory_ptr))));
}
// One of the linked object files has a reference to the __heap_base symbol.
// We must set its virtual address so it can be used in relocations.
if (wasm.findGlobalSymbol("__heap_base")) |loc| {
const symbol = loc.getSymbol(wasm);
symbol.virtual_address = @intCast(heap_alignment.forward(memory_ptr));
}
// Setup the max amount of pages
// For now we only support wasm32 by setting the maximum allowed memory size 2^32-1
const max_memory_allowed: u64 = (1 << 32) - 1;
if (wasm.base.options.initial_memory) |initial_memory| {
if (!std.mem.isAlignedGeneric(u64, initial_memory, page_size)) {
log.err("Initial memory must be {d}-byte aligned", .{page_size});
return error.MissAlignment;
}
if (memory_ptr > initial_memory) {
log.err("Initial memory too small, must be at least {d} bytes", .{memory_ptr});
return error.MemoryTooSmall;
}
if (initial_memory > max_memory_allowed) {
log.err("Initial memory exceeds maximum memory {d}", .{max_memory_allowed});
return error.MemoryTooBig;
}
memory_ptr = initial_memory;
}
memory_ptr = mem.alignForward(u64, memory_ptr, std.wasm.page_size);
// In case we do not import memory, but define it ourselves,
// set the minimum amount of pages on the memory section.
wasm.memories.limits.min = @as(u32, @intCast(memory_ptr / page_size));
log.debug("Total memory pages: {d}", .{wasm.memories.limits.min});
if (wasm.findGlobalSymbol("__heap_end")) |loc| {
const symbol = loc.getSymbol(wasm);
symbol.virtual_address = @as(u32, @intCast(memory_ptr));
}
if (wasm.base.options.max_memory) |max_memory| {
if (!std.mem.isAlignedGeneric(u64, max_memory, page_size)) {
log.err("Maximum memory must be {d}-byte aligned", .{page_size});
return error.MissAlignment;
}
if (memory_ptr > max_memory) {
log.err("Maxmimum memory too small, must be at least {d} bytes", .{memory_ptr});
return error.MemoryTooSmall;
}
if (max_memory > max_memory_allowed) {
log.err("Maximum memory exceeds maxmium amount {d}", .{max_memory_allowed});
return error.MemoryTooBig;
}
wasm.memories.limits.max = @as(u32, @intCast(max_memory / page_size));
wasm.memories.limits.setFlag(.WASM_LIMITS_FLAG_HAS_MAX);
if (wasm.base.options.shared_memory) {
wasm.memories.limits.setFlag(.WASM_LIMITS_FLAG_IS_SHARED);
}
log.debug("Maximum memory pages: {?d}", .{wasm.memories.limits.max});
}
}
/// From a given object's index and the index of the segment, returns the corresponding
/// index of the segment within the final data section. When the segment does not yet
/// exist, a new one will be initialized and appended. The new index will be returned in that case.
pub fn getMatchingSegment(wasm: *Wasm, object_index: u16, relocatable_index: u32) !?u32 {
const object: Object = wasm.objects.items[object_index];
const relocatable_data = object.relocatable_data[relocatable_index];
const index = @as(u32, @intCast(wasm.segments.items.len));
switch (relocatable_data.type) {
.data => {
const segment_info = object.segment_info[relocatable_data.index];
const merge_segment = wasm.base.options.output_mode != .Obj;
const result = try wasm.data_segments.getOrPut(wasm.base.allocator, segment_info.outputName(merge_segment));
if (!result.found_existing) {
result.value_ptr.* = index;
var flags: u32 = 0;
if (wasm.base.options.shared_memory) {
flags |= @intFromEnum(Segment.Flag.WASM_DATA_SEGMENT_IS_PASSIVE);
}
try wasm.segments.append(wasm.base.allocator, .{
.alignment = .@"1",
.size = 0,
.offset = 0,
.flags = flags,
});
return index;
} else return result.value_ptr.*;
},
.code => return wasm.code_section_index orelse blk: {
wasm.code_section_index = index;
try wasm.appendDummySegment();
break :blk index;
},
.debug => {
const debug_name = object.getDebugName(relocatable_data);
if (mem.eql(u8, debug_name, ".debug_info")) {
return wasm.debug_info_index orelse blk: {
wasm.debug_info_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_line")) {
return wasm.debug_line_index orelse blk: {
wasm.debug_line_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_loc")) {
return wasm.debug_loc_index orelse blk: {
wasm.debug_loc_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_ranges")) {
return wasm.debug_line_index orelse blk: {
wasm.debug_ranges_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_pubnames")) {
return wasm.debug_pubnames_index orelse blk: {
wasm.debug_pubnames_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_pubtypes")) {
return wasm.debug_pubtypes_index orelse blk: {
wasm.debug_pubtypes_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_abbrev")) {
return wasm.debug_abbrev_index orelse blk: {
wasm.debug_abbrev_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else if (mem.eql(u8, debug_name, ".debug_str")) {
return wasm.debug_str_index orelse blk: {
wasm.debug_str_index = index;
try wasm.appendDummySegment();
break :blk index;
};
} else {
log.warn("found unknown debug section '{s}'", .{debug_name});
log.warn(" debug section will be skipped", .{});
return null;
}
},
}
}
/// Appends a new segment with default field values
fn appendDummySegment(wasm: *Wasm) !void {
try wasm.segments.append(wasm.base.allocator, .{
.alignment = .@"1",
.size = 0,
.offset = 0,
.flags = 0,
});
}
/// Returns the symbol index of the error name table.
///
/// When the symbol does not yet exist, it will create a new one instead.
pub fn getErrorTableSymbol(wasm: *Wasm) !u32 {
if (wasm.error_table_symbol) |symbol| {
return symbol;
}
// no error was referenced yet, so create a new symbol and atom for it
// and then return said symbol's index. The final table will be populated
// during `flush` when we know all possible error names.
const atom_index = try wasm.createAtom();
const atom = wasm.getAtomPtr(atom_index);
const slice_ty = Type.slice_const_u8_sentinel_0;
const mod = wasm.base.options.module.?;
atom.alignment = slice_ty.abiAlignment(mod);
const sym_index = atom.sym_index;
const sym_name = try wasm.string_table.put(wasm.base.allocator, "__zig_err_name_table");
const symbol = &wasm.symbols.items[sym_index];
symbol.* = .{
.name = sym_name,
.tag = .data,
.flags = 0,
.index = 0,
.virtual_address = undefined,
};
symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
try wasm.resolved_symbols.put(wasm.base.allocator, atom.symbolLoc(), {});
log.debug("Error name table was created with symbol index: ({d})", .{sym_index});
wasm.error_table_symbol = sym_index;
return sym_index;
}
/// Populates the error name table, when `error_table_symbol` is not null.
///
/// This creates a table that consists of pointers and length to each error name.
/// The table is what is being pointed to within the runtime bodies that are generated.
fn populateErrorNameTable(wasm: *Wasm) !void {
const symbol_index = wasm.error_table_symbol orelse return;
const atom_index = wasm.symbol_atom.get(.{ .file = null, .index = symbol_index }).?;
// Rather than creating a symbol for each individual error name,
// we create a symbol for the entire region of error names. We then calculate
// the pointers into the list using addends which are appended to the relocation.
const names_atom_index = try wasm.createAtom();
const names_atom = wasm.getAtomPtr(names_atom_index);
names_atom.alignment = .@"1";
const sym_name = try wasm.string_table.put(wasm.base.allocator, "__zig_err_names");
const names_symbol = &wasm.symbols.items[names_atom.sym_index];
names_symbol.* = .{
.name = sym_name,
.tag = .data,
.flags = 0,
.index = 0,
.virtual_address = undefined,
};
names_symbol.setFlag(.WASM_SYM_VISIBILITY_HIDDEN);
log.debug("Populating error names", .{});
// Addend for each relocation to the table
var addend: u32 = 0;
const mod = wasm.base.options.module.?;
for (mod.global_error_set.keys()) |error_name_nts| {
const atom = wasm.getAtomPtr(atom_index);
const error_name = mod.intern_pool.stringToSlice(error_name_nts);
const len = @as(u32, @intCast(error_name.len + 1)); // names are 0-termianted
const slice_ty = Type.slice_const_u8_sentinel_0;
const offset = @as(u32, @intCast(atom.code.items.len));
// first we create the data for the slice of the name
try atom.code.appendNTimes(wasm.base.allocator, 0, 4); // ptr to name, will be relocated
try atom.code.writer(wasm.base.allocator).writeInt(u32, len - 1, .little);
// create relocation to the error name
try atom.relocs.append(wasm.base.allocator, .{
.index = names_atom.sym_index,
.relocation_type = .R_WASM_MEMORY_ADDR_I32,
.offset = offset,
.addend = @as(i32, @intCast(addend)),
});
atom.size += @as(u32, @intCast(slice_ty.abiSize(mod)));
addend += len;
// as we updated the error name table, we now store the actual name within the names atom
try names_atom.code.ensureUnusedCapacity(wasm.base.allocator, len);
names_atom.code.appendSliceAssumeCapacity(error_name);
names_atom.code.appendAssumeCapacity(0);
log.debug("Populated error name: '{s}'", .{error_name});
}
names_atom.size = addend;
const name_loc = names_atom.symbolLoc();
try wasm.resolved_symbols.put(wasm.base.allocator, name_loc, {});
try wasm.symbol_atom.put(wasm.base.allocator, name_loc, names_atom_index);
// link the atoms with the rest of the binary so they can be allocated
// and relocations will be performed.
try wasm.parseAtom(atom_index, .{ .data = .read_only });
try wasm.parseAtom(names_atom_index, .{ .data = .read_only });
}
/// From a given index variable, creates a new debug section.
/// This initializes the index, appends a new segment,
/// and finally, creates a managed `Atom`.
pub fn createDebugSectionForIndex(wasm: *Wasm, index: *?u32, name: []const u8) !Atom.Index {
const new_index = @as(u32, @intCast(wasm.segments.items.len));
index.* = new_index;
try wasm.appendDummySegment();
const atom_index = try wasm.createAtom();
const atom = wasm.getAtomPtr(atom_index);
wasm.symbols.items[atom.sym_index] = .{
.tag = .section,
.name = try wasm.string_table.put(wasm.base.allocator, name),
.index = 0,
.flags = @intFromEnum(Symbol.Flag.WASM_SYM_BINDING_LOCAL),
};
atom.alignment = .@"1"; // debug sections are always 1-byte-aligned
return atom_index;
}
fn resetState(wasm: *Wasm) void {
for (wasm.segment_info.values()) |segment_info| {
wasm.base.allocator.free(segment_info.name);
}
var atom_it = wasm.decls.valueIterator();
while (atom_it.next()) |atom_index| {
const atom = wasm.getAtomPtr(atom_index.*);
atom.next = null;
atom.prev = null;
for (atom.locals.items) |local_atom_index| {
const local_atom = wasm.getAtomPtr(local_atom_index);
local_atom.next = null;
local_atom.prev = null;
}
}
wasm.functions.clearRetainingCapacity();
wasm.exports.clearRetainingCapacity();
wasm.segments.clearRetainingCapacity();
wasm.segment_info.clearRetainingCapacity();
wasm.data_segments.clearRetainingCapacity();
wasm.atoms.clearRetainingCapacity();
wasm.symbol_atom.clearRetainingCapacity();
wasm.code_section_index = null;
wasm.debug_info_index = null;
wasm.debug_line_index = null;
wasm.debug_loc_index = null;
wasm.debug_str_index = null;
wasm.debug_ranges_index = null;
wasm.debug_abbrev_index = null;
wasm.debug_pubnames_index = null;
wasm.debug_pubtypes_index = null;
}
pub fn flush(wasm: *Wasm, comp: *Compilation, prog_node: *std.Progress.Node) link.File.FlushError!void {
if (wasm.base.options.emit == null) {
if (wasm.llvm_object) |llvm_object| {
return try llvm_object.flushModule(comp, prog_node);
}
return;
}
if (build_options.have_llvm and wasm.base.options.use_lld) {
return wasm.linkWithLLD(comp, prog_node);
} else if (wasm.base.options.use_llvm and !wasm.base.options.use_lld) {
return wasm.linkWithZld(comp, prog_node);
} else {
return wasm.flushModule(comp, prog_node);
}
}
/// Uses the in-house linker to link one or multiple object -and archive files into a WebAssembly binary.
fn linkWithZld(wasm: *Wasm, comp: *Compilation, prog_node: *std.Progress.Node) link.File.FlushError!void {
const tracy = trace(@src());
defer tracy.end();
const gpa = wasm.base.allocator;
const options = wasm.base.options;
// Used for all temporary memory allocated during flushin
var arena_instance = std.heap.ArenaAllocator.init(gpa);
defer arena_instance.deinit();
const arena = arena_instance.allocator();
const directory = options.emit.?.directory; // Just an alias to make it shorter to type.
const full_out_path = try directory.join(arena, &[_][]const u8{options.emit.?.sub_path});
// If there is no Zig code to compile, then we should skip flushing the output file because it
// will not be part of the linker line anyway.
const module_obj_path: ?[]const u8 = if (options.module != null) blk: {
assert(options.use_llvm); // `linkWithZld` should never be called when the Wasm backend is used
try wasm.flushModule(comp, prog_node);
if (fs.path.dirname(full_out_path)) |dirname| {
break :blk try fs.path.join(arena, &.{ dirname, wasm.base.intermediary_basename.? });
} else {
break :blk wasm.base.intermediary_basename.?;
}
} else null;
var sub_prog_node = prog_node.start("Wasm Flush", 0);
sub_prog_node.activate();
defer sub_prog_node.end();
const compiler_rt_path: ?[]const u8 = blk: {
if (comp.compiler_rt_obj) |obj| break :blk obj.full_object_path;
if (comp.compiler_rt_lib) |lib| break :blk lib.full_object_path;
break :blk null;
};
const id_symlink_basename = "zld.id";
var man: Cache.Manifest = undefined;
defer if (!options.disable_lld_caching) man.deinit();
var digest: [Cache.hex_digest_len]u8 = undefined;
// NOTE: The following section must be maintained to be equal
// as the section defined in `linkWithLLD`
if (!options.disable_lld_caching) {
man = comp.cache_parent.obtain();
// We are about to obtain this lock, so here we give other processes a chance first.
wasm.base.releaseLock();
comptime assert(Compilation.link_hash_implementation_version == 10);
for (options.objects) |obj| {
_ = try man.addFile(obj.path, null);
man.hash.add(obj.must_link);
}
for (comp.c_object_table.keys()) |key| {
_ = try man.addFile(key.status.success.object_path, null);
}
try man.addOptionalFile(module_obj_path);
try man.addOptionalFile(compiler_rt_path);
man.hash.addOptionalBytes(options.entry);
man.hash.addOptional(options.stack_size_override);
man.hash.add(wasm.base.options.build_id);
man.hash.add(options.import_memory);
man.hash.add(options.import_table);
man.hash.add(options.export_table);
man.hash.addOptional(options.initial_memory);
man.hash.addOptional(options.max_memory);
man.hash.add(options.shared_memory);
man.hash.addOptional(options.global_base);
man.hash.add(options.export_symbol_names.len);
// strip does not need to go into the linker hash because it is part of the hash namespace
for (options.export_symbol_names) |symbol_name| {
man.hash.addBytes(symbol_name);
}
// We don't actually care whether it's a cache hit or miss; we just need the digest and the lock.
_ = try man.hit();
digest = man.final();
var prev_digest_buf: [digest.len]u8 = undefined;
const prev_digest: []u8 = Cache.readSmallFile(
directory.handle,
id_symlink_basename,
&prev_digest_buf,
) catch |err| blk: {
log.debug("WASM LLD new_digest={s} error: {s}", .{ std.fmt.fmtSliceHexLower(&digest), @errorName(err) });
// Handle this as a cache miss.
break :blk prev_digest_buf[0..0];
};
if (mem.eql(u8, prev_digest, &digest)) {
log.debug("WASM LLD digest={s} match - skipping invocation", .{std.fmt.fmtSliceHexLower(&digest)});
// Hot diggity dog! The output binary is already there.
wasm.base.lock = man.toOwnedLock();
return;
}
log.debug("WASM LLD prev_digest={s} new_digest={s}", .{ std.fmt.fmtSliceHexLower(prev_digest), std.fmt.fmtSliceHexLower(&digest) });
// We are about to change the output file to be different, so we invalidate the build hash now.
directory.handle.deleteFile(id_symlink_basename) catch |err| switch (err) {
error.FileNotFound => {},
else => |e| return e,
};
}
// Positional arguments to the linker such as object files and static archives.
var positionals = std.ArrayList([]const u8).init(arena);
try positionals.ensureUnusedCapacity(options.objects.len);
// When the target os is WASI, we allow linking with WASI-LIBC
if (options.target.os.tag == .wasi) {
const is_exe_or_dyn_lib = wasm.base.options.output_mode == .Exe or
(wasm.base.options.output_mode == .Lib and wasm.base.options.link_mode == .Dynamic);
if (is_exe_or_dyn_lib) {
const wasi_emulated_libs = wasm.base.options.wasi_emulated_libs;
for (wasi_emulated_libs) |crt_file| {
try positionals.append(try comp.get_libc_crt_file(
arena,
wasi_libc.emulatedLibCRFileLibName(crt_file),
));
}
if (wasm.base.options.link_libc) {
try positionals.append(try comp.get_libc_crt_file(
arena,
wasi_libc.execModelCrtFileFullName(wasm.base.options.wasi_exec_model),
));
try positionals.append(try comp.get_libc_crt_file(arena, "libc.a"));
}
if (wasm.base.options.link_libcpp) {
try positionals.append(comp.libcxx_static_lib.?.full_object_path);
try positionals.append(comp.libcxxabi_static_lib.?.full_object_path);
}
}
}
if (module_obj_path) |path| {
try positionals.append(path);
}
for (options.objects) |object| {
try positionals.append(object.path);
}
for (comp.c_object_table.keys()) |c_object| {
try positionals.append(c_object.status.success.object_path);
}
if (comp.compiler_rt_lib) |lib| try positionals.append(lib.full_object_path);
if (comp.compiler_rt_obj) |obj| try positionals.append(obj.full_object_path);
try wasm.parseInputFiles(positionals.items);
for (wasm.objects.items, 0..) |_, object_index| {
try wasm.resolveSymbolsInObject(@as(u16, @intCast(object_index)));
}
var emit_features_count: u32 = 0;
var enabled_features: [@typeInfo(types.Feature.Tag).Enum.fields.len]bool = undefined;
try wasm.validateFeatures(&enabled_features, &emit_features_count);
try wasm.resolveSymbolsInArchives();
try wasm.resolveLazySymbols();
try wasm.checkUndefinedSymbols();
try wasm.setupInitFunctions();
try wasm.setupStart();
try wasm.setupImports();
for (wasm.objects.items, 0..) |*object, object_index| {
try object.parseIntoAtoms(gpa, @as(u16, @intCast(object_index)), wasm);
}
try wasm.allocateAtoms();
try wasm.setupMemory();
wasm.allocateVirtualAddresses();
wasm.mapFunctionTable();
try wasm.mergeSections();
try wasm.mergeTypes();
try wasm.initializeCallCtorsFunction();
try wasm.setupInitMemoryFunction();
try wasm.setupTLSRelocationsFunction();
try wasm.initializeTLSFunction();
try wasm.setupStartSection();
try wasm.setupExports();
try wasm.writeToFile(enabled_features, emit_features_count, arena);
if (!wasm.base.options.disable_lld_caching) {
// Update the file with the digest. If it fails we can continue; it only
// means that the next invocation will have an unnecessary cache miss.
Cache.writeSmallFile(directory.handle, id_symlink_basename, &digest) catch |err| {
log.warn("failed to save linking hash digest symlink: {s}", .{@errorName(err)});
};
// Again failure here only means an unnecessary cache miss.
man.writeManifest() catch |err| {
log.warn("failed to write cache manifest when linking: {s}", .{@errorName(err)});
};
// We hang on to this lock so that the output file path can be used without
// other processes clobbering it.
wasm.base.lock = man.toOwnedLock();
}
}
pub fn flushModule(wasm: *Wasm, comp: *Compilation, prog_node: *std.Progress.Node) link.File.FlushError!void {
const tracy = trace(@src());
defer tracy.end();
if (wasm.llvm_object) |llvm_object| {
return try llvm_object.flushModule(comp, prog_node);
}
var sub_prog_node = prog_node.start("Wasm Flush", 0);
sub_prog_node.activate();
defer sub_prog_node.end();
// ensure the error names table is populated when an error name is referenced
try wasm.populateErrorNameTable();
// Used for all temporary memory allocated during flushin
var arena_instance = std.heap.ArenaAllocator.init(wasm.base.allocator);
defer arena_instance.deinit();
const arena = arena_instance.allocator();
// Positional arguments to the linker such as object files and static archives.
var positionals = std.ArrayList([]const u8).init(arena);
try positionals.ensureUnusedCapacity(wasm.base.options.objects.len);
for (wasm.base.options.objects) |object| {
positionals.appendAssumeCapacity(object.path);
}
for (comp.c_object_table.keys()) |c_object| {
try positionals.append(c_object.status.success.object_path);
}
if (comp.compiler_rt_lib) |lib| try positionals.append(lib.full_object_path);
if (comp.compiler_rt_obj) |obj| try positionals.append(obj.full_object_path);
try wasm.parseInputFiles(positionals.items);
for (wasm.objects.items, 0..) |_, object_index| {
try wasm.resolveSymbolsInObject(@as(u16, @intCast(object_index)));
}
var emit_features_count: u32 = 0;
var enabled_features: [@typeInfo(types.Feature.Tag).Enum.fields.len]bool = undefined;
try wasm.validateFeatures(&enabled_features, &emit_features_count);
try wasm.resolveSymbolsInArchives();
try wasm.resolveLazySymbols();
try wasm.checkUndefinedSymbols();
// When we finish/error we reset the state of the linker
// So we can rebuild the binary file on each incremental update
defer wasm.resetState();
try wasm.setupInitFunctions();
try wasm.setupErrorsLen();
try wasm.setupStart();
try wasm.setupImports();
if (wasm.base.options.module) |mod| {
var decl_it = wasm.decls.iterator();
while (decl_it.next()) |entry| {
const decl = mod.declPtr(entry.key_ptr.*);
if (decl.isExtern(mod)) continue;
const atom_index = entry.value_ptr.*;
const atom = wasm.getAtomPtr(atom_index);
if (decl.ty.zigTypeTag(mod) == .Fn) {
try wasm.parseAtom(atom_index, .function);
} else if (decl.getOwnedVariable(mod)) |variable| {
if (variable.is_const) {
try wasm.parseAtom(atom_index, .{ .data = .read_only });
} else if (variable.init.toValue().isUndefDeep(mod)) {
// for safe build modes, we store the atom in the data segment,
// whereas for unsafe build modes we store it in bss.
const is_initialized = wasm.base.options.optimize_mode == .Debug or
wasm.base.options.optimize_mode == .ReleaseSafe;
try wasm.parseAtom(atom_index, .{ .data = if (is_initialized) .initialized else .uninitialized });
} else {
// when the decl is all zeroes, we store the atom in the bss segment,
// in all other cases it will be in the data segment.
const is_zeroes = for (atom.code.items) |byte| {
if (byte != 0) break false;
} else true;
try wasm.parseAtom(atom_index, .{ .data = if (is_zeroes) .uninitialized else .initialized });
}
} else {
try wasm.parseAtom(atom_index, .{ .data = .read_only });
}
// also parse atoms for a decl's locals
for (atom.locals.items) |local_atom_index| {
try wasm.parseAtom(local_atom_index, .{ .data = .read_only });
}
}
// parse anonymous declarations
for (wasm.anon_decls.keys(), wasm.anon_decls.values()) |decl_val, atom_index| {
const ty = mod.intern_pool.typeOf(decl_val).toType();
if (ty.zigTypeTag(mod) == .Fn) {
try wasm.parseAtom(atom_index, .function);
} else {
try wasm.parseAtom(atom_index, .{ .data = .read_only });
}
}
// also parse any backend-generated functions
for (wasm.synthetic_functions.items) |atom_index| {
try wasm.parseAtom(atom_index, .function);
}
if (wasm.dwarf) |*dwarf| {
try dwarf.flushModule(wasm.base.options.module.?);
}
}
for (wasm.objects.items, 0..) |*object, object_index| {
try object.parseIntoAtoms(wasm.base.allocator, @as(u16, @intCast(object_index)), wasm);
}
try wasm.allocateAtoms();
try wasm.setupMemory();
wasm.allocateVirtualAddresses();
wasm.mapFunctionTable();
try wasm.mergeSections();
try wasm.mergeTypes();
try wasm.initializeCallCtorsFunction();
try wasm.setupInitMemoryFunction();
try wasm.setupTLSRelocationsFunction();
try wasm.initializeTLSFunction();
try wasm.setupStartSection();
try wasm.setupExports();
try wasm.writeToFile(enabled_features, emit_features_count, arena);
}
/// Writes the WebAssembly in-memory module to the file
fn writeToFile(
wasm: *Wasm,
enabled_features: [@typeInfo(types.Feature.Tag).Enum.fields.len]bool,
feature_count: u32,
arena: Allocator,
) !void {
// Size of each section header
const header_size = 5 + 1;
// The amount of sections that will be written
var section_count: u32 = 0;
// Index of the code section. Used to tell relocation table where the section lives.
var code_section_index: ?u32 = null;
// Index of the data section. Used to tell relocation table where the section lives.
var data_section_index: ?u32 = null;
const is_obj = wasm.base.options.output_mode == .Obj or (!wasm.base.options.use_llvm and wasm.base.options.use_lld);
var binary_bytes = std.ArrayList(u8).init(wasm.base.allocator);
defer binary_bytes.deinit();
const binary_writer = binary_bytes.writer();
// We write the magic bytes at the end so they will only be written
// if everything succeeded as expected. So populate with 0's for now.
try binary_writer.writeAll(&[_]u8{0} ** 8);
// (Re)set file pointer to 0
try wasm.base.file.?.setEndPos(0);
try wasm.base.file.?.seekTo(0);
// Type section
if (wasm.func_types.items.len != 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
log.debug("Writing type section. Count: ({d})", .{wasm.func_types.items.len});
for (wasm.func_types.items) |func_type| {
try leb.writeULEB128(binary_writer, std.wasm.function_type);
try leb.writeULEB128(binary_writer, @as(u32, @intCast(func_type.params.len)));
for (func_type.params) |param_ty| {
try leb.writeULEB128(binary_writer, std.wasm.valtype(param_ty));
}
try leb.writeULEB128(binary_writer, @as(u32, @intCast(func_type.returns.len)));
for (func_type.returns) |ret_ty| {
try leb.writeULEB128(binary_writer, std.wasm.valtype(ret_ty));
}
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.type,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.func_types.items.len)),
);
section_count += 1;
}
// Import section
const import_memory = wasm.base.options.import_memory or is_obj;
const export_memory = wasm.base.options.export_memory;
if (wasm.imports.count() != 0 or import_memory) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
var it = wasm.imports.iterator();
while (it.next()) |entry| {
assert(entry.key_ptr.*.getSymbol(wasm).isUndefined());
const import = entry.value_ptr.*;
try wasm.emitImport(binary_writer, import);
}
if (import_memory) {
const mem_name = if (is_obj) "__linear_memory" else "memory";
const mem_imp: types.Import = .{
.module_name = try wasm.string_table.put(wasm.base.allocator, wasm.host_name),
.name = try wasm.string_table.put(wasm.base.allocator, mem_name),
.kind = .{ .memory = wasm.memories.limits },
};
try wasm.emitImport(binary_writer, mem_imp);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.import,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.imports.count() + @intFromBool(import_memory))),
);
section_count += 1;
}
// Function section
if (wasm.functions.count() != 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
for (wasm.functions.values()) |function| {
try leb.writeULEB128(binary_writer, function.type_index);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.function,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.functions.count())),
);
section_count += 1;
}
// Table section
if (wasm.tables.items.len > 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
for (wasm.tables.items) |table| {
try leb.writeULEB128(binary_writer, std.wasm.reftype(table.reftype));
try emitLimits(binary_writer, table.limits);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.table,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.tables.items.len)),
);
section_count += 1;
}
// Memory section
if (!import_memory) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
try emitLimits(binary_writer, wasm.memories.limits);
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.memory,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, 1), // wasm currently only supports 1 linear memory segment
);
section_count += 1;
}
// Global section (used to emit stack pointer)
if (wasm.wasm_globals.items.len > 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
for (wasm.wasm_globals.items) |global| {
try binary_writer.writeByte(std.wasm.valtype(global.global_type.valtype));
try binary_writer.writeByte(@intFromBool(global.global_type.mutable));
try emitInit(binary_writer, global.init);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.global,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.wasm_globals.items.len)),
);
section_count += 1;
}
// Export section
if (wasm.exports.items.len != 0 or export_memory) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
for (wasm.exports.items) |exp| {
const name = wasm.string_table.get(exp.name);
try leb.writeULEB128(binary_writer, @as(u32, @intCast(name.len)));
try binary_writer.writeAll(name);
try leb.writeULEB128(binary_writer, @intFromEnum(exp.kind));
try leb.writeULEB128(binary_writer, exp.index);
}
if (export_memory) {
try leb.writeULEB128(binary_writer, @as(u32, @intCast("memory".len)));
try binary_writer.writeAll("memory");
try binary_writer.writeByte(std.wasm.externalKind(.memory));
try leb.writeULEB128(binary_writer, @as(u32, 0));
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.@"export",
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(wasm.exports.items.len)) + @intFromBool(export_memory),
);
section_count += 1;
}
if (wasm.entry) |entry_index| {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.start,
@intCast(binary_bytes.items.len - header_offset - header_size),
entry_index,
);
}
// element section (function table)
if (wasm.function_table.count() > 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
const table_loc = wasm.findGlobalSymbol("__indirect_function_table").?;
const table_sym = table_loc.getSymbol(wasm);
var flags: u32 = if (table_sym.index == 0) 0x0 else 0x02; // passive with implicit 0-index table or set table index manually
try leb.writeULEB128(binary_writer, flags);
if (flags == 0x02) {
try leb.writeULEB128(binary_writer, table_sym.index);
}
try emitInit(binary_writer, .{ .i32_const = 1 }); // We start at index 1, so unresolved function pointers are invalid
if (flags == 0x02) {
try leb.writeULEB128(binary_writer, @as(u8, 0)); // represents funcref
}
try leb.writeULEB128(binary_writer, @as(u32, @intCast(wasm.function_table.count())));
var symbol_it = wasm.function_table.keyIterator();
while (symbol_it.next()) |symbol_loc_ptr| {
try leb.writeULEB128(binary_writer, symbol_loc_ptr.*.getSymbol(wasm).index);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.element,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, 1),
);
section_count += 1;
}
// When the shared-memory option is enabled, we *must* emit the 'data count' section.
const data_segments_count = wasm.data_segments.count() - @intFromBool(wasm.data_segments.contains(".bss") and !import_memory);
if (data_segments_count != 0 and wasm.base.options.shared_memory) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.data_count,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(data_segments_count)),
);
}
// Code section
var code_section_size: u32 = 0;
if (wasm.code_section_index) |code_index| {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
var atom_index = wasm.atoms.get(code_index).?;
// The code section must be sorted in line with the function order.
var sorted_atoms = try std.ArrayList(*Atom).initCapacity(wasm.base.allocator, wasm.functions.count());
defer sorted_atoms.deinit();
while (true) {
var atom = wasm.getAtomPtr(atom_index);
if (wasm.resolved_symbols.contains(atom.symbolLoc())) {
if (!is_obj) {
atom.resolveRelocs(wasm);
}
sorted_atoms.appendAssumeCapacity(atom);
}
// atom = if (atom.prev) |prev| wasm.getAtomPtr(prev) else break;
atom_index = atom.prev orelse break;
}
const atom_sort_fn = struct {
fn sort(ctx: *const Wasm, lhs: *const Atom, rhs: *const Atom) bool {
const lhs_sym = lhs.symbolLoc().getSymbol(ctx);
const rhs_sym = rhs.symbolLoc().getSymbol(ctx);
return lhs_sym.index < rhs_sym.index;
}
}.sort;
mem.sort(*Atom, sorted_atoms.items, wasm, atom_sort_fn);
for (sorted_atoms.items) |sorted_atom| {
try leb.writeULEB128(binary_writer, sorted_atom.size);
try binary_writer.writeAll(sorted_atom.code.items);
}
code_section_size = @as(u32, @intCast(binary_bytes.items.len - header_offset - header_size));
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.code,
code_section_size,
@as(u32, @intCast(wasm.functions.count())),
);
code_section_index = section_count;
section_count += 1;
}
// Data section
if (data_segments_count != 0) {
const header_offset = try reserveVecSectionHeader(&binary_bytes);
var it = wasm.data_segments.iterator();
var segment_count: u32 = 0;
while (it.next()) |entry| {
// do not output 'bss' section unless we import memory and therefore
// want to guarantee the data is zero initialized
if (!import_memory and std.mem.eql(u8, entry.key_ptr.*, ".bss")) continue;
const segment_index = entry.value_ptr.*;
const segment = wasm.segments.items[segment_index];
if (segment.size == 0) continue; // do not emit empty segments
segment_count += 1;
var atom_index = wasm.atoms.get(segment_index).?;
try leb.writeULEB128(binary_writer, segment.flags);
if (segment.flags & @intFromEnum(Wasm.Segment.Flag.WASM_DATA_SEGMENT_HAS_MEMINDEX) != 0) {
try leb.writeULEB128(binary_writer, @as(u32, 0)); // memory is always index 0 as we only have 1 memory entry
}
// when a segment is passive, it's initialized during runtime.
if (!segment.isPassive()) {
try emitInit(binary_writer, .{ .i32_const = @as(i32, @bitCast(segment.offset)) });
}
// offset into data section
try leb.writeULEB128(binary_writer, segment.size);
// fill in the offset table and the data segments
var current_offset: u32 = 0;
while (true) {
const atom = wasm.getAtomPtr(atom_index);
if (!is_obj) {
atom.resolveRelocs(wasm);
}
// Pad with zeroes to ensure all segments are aligned
if (current_offset != atom.offset) {
const diff = atom.offset - current_offset;
try binary_writer.writeByteNTimes(0, diff);
current_offset += diff;
}
assert(current_offset == atom.offset);
assert(atom.code.items.len == atom.size);
try binary_writer.writeAll(atom.code.items);
current_offset += atom.size;
if (atom.prev) |prev| {
atom_index = prev;
} else {
// also pad with zeroes when last atom to ensure
// segments are aligned.
if (current_offset != segment.size) {
try binary_writer.writeByteNTimes(0, segment.size - current_offset);
current_offset += segment.size - current_offset;
}
break;
}
}
assert(current_offset == segment.size);
}
try writeVecSectionHeader(
binary_bytes.items,
header_offset,
.data,
@as(u32, @intCast(binary_bytes.items.len - header_offset - header_size)),
@as(u32, @intCast(segment_count)),
);
data_section_index = section_count;
section_count += 1;
}
if (is_obj) {
// relocations need to point to the index of a symbol in the final symbol table. To save memory,
// we never store all symbols in a single table, but store a location reference instead.
// This means that for a relocatable object file, we need to generate one and provide it to the relocation sections.
var symbol_table = std.AutoArrayHashMap(SymbolLoc, u32).init(arena);
try wasm.emitLinkSection(&binary_bytes, &symbol_table);
if (code_section_index) |code_index| {
try wasm.emitCodeRelocations(&binary_bytes, code_index, symbol_table);
}
if (data_section_index) |data_index| {
try wasm.emitDataRelocations(&binary_bytes, data_index, symbol_table);
}
} else if (!wasm.base.options.strip) {
try wasm.emitNameSection(&binary_bytes, arena);
}
if (!wasm.base.options.strip) {
// The build id must be computed on the main sections only,
// so we have to do it now, before the debug sections.
switch (wasm.base.options.build_id) {
.none => {},
.fast => {
var id: [16]u8 = undefined;
std.crypto.hash.sha3.TurboShake128(null).hash(binary_bytes.items, &id, .{});
var uuid: [36]u8 = undefined;
_ = try std.fmt.bufPrint(&uuid, "{s}-{s}-{s}-{s}-{s}", .{
std.fmt.fmtSliceHexLower(id[0..4]),
std.fmt.fmtSliceHexLower(id[4..6]),
std.fmt.fmtSliceHexLower(id[6..8]),
std.fmt.fmtSliceHexLower(id[8..10]),
std.fmt.fmtSliceHexLower(id[10..]),
});
try emitBuildIdSection(&binary_bytes, &uuid);
},
.hexstring => |hs| {
var buffer: [32 * 2]u8 = undefined;
const str = std.fmt.bufPrint(&buffer, "{s}", .{
std.fmt.fmtSliceHexLower(hs.toSlice()),
}) catch unreachable;
try emitBuildIdSection(&binary_bytes, str);
},
else => |mode| log.err("build-id '{s}' is not supported for WASM", .{@tagName(mode)}),
}
// if (wasm.dwarf) |*dwarf| {
// const mod = wasm.base.options.module.?;
// try dwarf.writeDbgAbbrev();
// // for debug info and ranges, the address is always 0,
// // as locations are always offsets relative to 'code' section.
// try dwarf.writeDbgInfoHeader(mod, 0, code_section_size);
// try dwarf.writeDbgAranges(0, code_section_size);
// try dwarf.writeDbgLineHeader();
// }
var debug_bytes = std.ArrayList(u8).init(wasm.base.allocator);
defer debug_bytes.deinit();
const DebugSection = struct {
name: []const u8,
index: ?u32,
};
const debug_sections: []const DebugSection = &.{
.{ .name = ".debug_info", .index = wasm.debug_info_index },
.{ .name = ".debug_pubtypes", .index = wasm.debug_pubtypes_index },
.{ .name = ".debug_abbrev", .index = wasm.debug_abbrev_index },
.{ .name = ".debug_line", .index = wasm.debug_line_index },
.{ .name = ".debug_str", .index = wasm.debug_str_index },
.{ .name = ".debug_pubnames", .index = wasm.debug_pubnames_index },
.{ .name = ".debug_loc", .index = wasm.debug_loc_index },
.{ .name = ".debug_ranges", .index = wasm.debug_ranges_index },
};
for (debug_sections) |item| {
if (item.index) |index| {
var atom = wasm.getAtomPtr(wasm.atoms.get(index).?);
while (true) {
atom.resolveRelocs(wasm);
try debug_bytes.appendSlice(atom.code.items);
atom = if (atom.prev) |prev| wasm.getAtomPtr(prev) else break;
}
try emitDebugSection(&binary_bytes, debug_bytes.items, item.name);
debug_bytes.clearRetainingCapacity();
}
}
try emitProducerSection(&binary_bytes);
if (feature_count > 0) {
try emitFeaturesSection(&binary_bytes, &enabled_features, feature_count);
}
}
// Only when writing all sections executed properly we write the magic
// bytes. This allows us to easily detect what went wrong while generating
// the final binary.
{
const src = std.wasm.magic ++ std.wasm.version;
binary_bytes.items[0..src.len].* = src;
}
// finally, write the entire binary into the file.
var iovec = [_]std.os.iovec_const{.{
.iov_base = binary_bytes.items.ptr,
.iov_len = binary_bytes.items.len,
}};
try wasm.base.file.?.writevAll(&iovec);
}
fn emitDebugSection(binary_bytes: *std.ArrayList(u8), data: []const u8, name: []const u8) !void {
if (data.len == 0) return;
const header_offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
try leb.writeULEB128(writer, @as(u32, @intCast(name.len)));
try writer.writeAll(name);
const start = binary_bytes.items.len - header_offset;
log.debug("Emit debug section: '{s}' start=0x{x:0>8} end=0x{x:0>8}", .{ name, start, start + data.len });
try writer.writeAll(data);
try writeCustomSectionHeader(
binary_bytes.items,
header_offset,
@as(u32, @intCast(binary_bytes.items.len - header_offset - 6)),
);
}
fn emitProducerSection(binary_bytes: *std.ArrayList(u8)) !void {
const header_offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
const producers = "producers";
try leb.writeULEB128(writer, @as(u32, @intCast(producers.len)));
try writer.writeAll(producers);
try leb.writeULEB128(writer, @as(u32, 2)); // 2 fields: Language + processed-by
// used for the Zig version
var version_buf: [100]u8 = undefined;
const version = try std.fmt.bufPrint(&version_buf, "{}", .{build_options.semver});
// language field
{
const language = "language";
try leb.writeULEB128(writer, @as(u32, @intCast(language.len)));
try writer.writeAll(language);
// field_value_count (TODO: Parse object files for producer sections to detect their language)
try leb.writeULEB128(writer, @as(u32, 1));
// versioned name
{
try leb.writeULEB128(writer, @as(u32, 3)); // len of "Zig"
try writer.writeAll("Zig");
try leb.writeULEB128(writer, @as(u32, @intCast(version.len)));
try writer.writeAll(version);
}
}
// processed-by field
{
const processed_by = "processed-by";
try leb.writeULEB128(writer, @as(u32, @intCast(processed_by.len)));
try writer.writeAll(processed_by);
// field_value_count (TODO: Parse object files for producer sections to detect other used tools)
try leb.writeULEB128(writer, @as(u32, 1));
// versioned name
{
try leb.writeULEB128(writer, @as(u32, 3)); // len of "Zig"
try writer.writeAll("Zig");
try leb.writeULEB128(writer, @as(u32, @intCast(version.len)));
try writer.writeAll(version);
}
}
try writeCustomSectionHeader(
binary_bytes.items,
header_offset,
@as(u32, @intCast(binary_bytes.items.len - header_offset - 6)),
);
}
fn emitBuildIdSection(binary_bytes: *std.ArrayList(u8), build_id: []const u8) !void {
const header_offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
const hdr_build_id = "build_id";
try leb.writeULEB128(writer, @as(u32, @intCast(hdr_build_id.len)));
try writer.writeAll(hdr_build_id);
try leb.writeULEB128(writer, @as(u32, 1));
try leb.writeULEB128(writer, @as(u32, @intCast(build_id.len)));
try writer.writeAll(build_id);
try writeCustomSectionHeader(
binary_bytes.items,
header_offset,
@as(u32, @intCast(binary_bytes.items.len - header_offset - 6)),
);
}
fn emitFeaturesSection(binary_bytes: *std.ArrayList(u8), enabled_features: []const bool, features_count: u32) !void {
const header_offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
const target_features = "target_features";
try leb.writeULEB128(writer, @as(u32, @intCast(target_features.len)));
try writer.writeAll(target_features);
try leb.writeULEB128(writer, features_count);
for (enabled_features, 0..) |enabled, feature_index| {
if (enabled) {
const feature: types.Feature = .{ .prefix = .used, .tag = @as(types.Feature.Tag, @enumFromInt(feature_index)) };
try leb.writeULEB128(writer, @intFromEnum(feature.prefix));
var buf: [100]u8 = undefined;
const string = try std.fmt.bufPrint(&buf, "{}", .{feature.tag});
try leb.writeULEB128(writer, @as(u32, @intCast(string.len)));
try writer.writeAll(string);
}
}
try writeCustomSectionHeader(
binary_bytes.items,
header_offset,
@as(u32, @intCast(binary_bytes.items.len - header_offset - 6)),
);
}
fn emitNameSection(wasm: *Wasm, binary_bytes: *std.ArrayList(u8), arena: std.mem.Allocator) !void {
const Name = struct {
index: u32,
name: []const u8,
fn lessThan(context: void, lhs: @This(), rhs: @This()) bool {
_ = context;
return lhs.index < rhs.index;
}
};
// we must de-duplicate symbols that point to the same function
var funcs = std.AutoArrayHashMap(u32, Name).init(arena);
try funcs.ensureUnusedCapacity(wasm.functions.count() + wasm.imported_functions_count);
var globals = try std.ArrayList(Name).initCapacity(arena, wasm.wasm_globals.items.len + wasm.imported_globals_count);
var segments = try std.ArrayList(Name).initCapacity(arena, wasm.data_segments.count());
for (wasm.resolved_symbols.keys()) |sym_loc| {
const symbol = sym_loc.getSymbol(wasm).*;
const name = sym_loc.getName(wasm);
switch (symbol.tag) {
.function => {
const gop = funcs.getOrPutAssumeCapacity(symbol.index);
if (!gop.found_existing) {
gop.value_ptr.* = .{ .index = symbol.index, .name = name };
}
},
.global => globals.appendAssumeCapacity(.{ .index = symbol.index, .name = name }),
else => {},
}
}
// data segments are already 'ordered'
var data_segment_index: u32 = 0;
for (wasm.data_segments.keys()) |key| {
// bss section is not emitted when this condition holds true, so we also
// do not output a name for it.
if (!wasm.base.options.import_memory and std.mem.eql(u8, key, ".bss")) continue;
segments.appendAssumeCapacity(.{ .index = data_segment_index, .name = key });
data_segment_index += 1;
}
mem.sort(Name, funcs.values(), {}, Name.lessThan);
mem.sort(Name, globals.items, {}, Name.lessThan);
const header_offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
try leb.writeULEB128(writer, @as(u32, @intCast("name".len)));
try writer.writeAll("name");
try wasm.emitNameSubsection(.function, funcs.values(), writer);
try wasm.emitNameSubsection(.global, globals.items, writer);
try wasm.emitNameSubsection(.data_segment, segments.items, writer);
try writeCustomSectionHeader(
binary_bytes.items,
header_offset,
@as(u32, @intCast(binary_bytes.items.len - header_offset - 6)),
);
}
fn emitNameSubsection(wasm: *Wasm, section_id: std.wasm.NameSubsection, names: anytype, writer: anytype) !void {
// We must emit subsection size, so first write to a temporary list
var section_list = std.ArrayList(u8).init(wasm.base.allocator);
defer section_list.deinit();
const sub_writer = section_list.writer();
try leb.writeULEB128(sub_writer, @as(u32, @intCast(names.len)));
for (names) |name| {
log.debug("Emit symbol '{s}' type({s})", .{ name.name, @tagName(section_id) });
try leb.writeULEB128(sub_writer, name.index);
try leb.writeULEB128(sub_writer, @as(u32, @intCast(name.name.len)));
try sub_writer.writeAll(name.name);
}
// From now, write to the actual writer
try leb.writeULEB128(writer, @intFromEnum(section_id));
try leb.writeULEB128(writer, @as(u32, @intCast(section_list.items.len)));
try writer.writeAll(section_list.items);
}
fn emitLimits(writer: anytype, limits: std.wasm.Limits) !void {
try writer.writeByte(limits.flags);
try leb.writeULEB128(writer, limits.min);
if (limits.hasFlag(.WASM_LIMITS_FLAG_HAS_MAX)) {
try leb.writeULEB128(writer, limits.max);
}
}
fn emitInit(writer: anytype, init_expr: std.wasm.InitExpression) !void {
switch (init_expr) {
.i32_const => |val| {
try writer.writeByte(std.wasm.opcode(.i32_const));
try leb.writeILEB128(writer, val);
},
.i64_const => |val| {
try writer.writeByte(std.wasm.opcode(.i64_const));
try leb.writeILEB128(writer, val);
},
.f32_const => |val| {
try writer.writeByte(std.wasm.opcode(.f32_const));
try writer.writeInt(u32, @bitCast(val), .little);
},
.f64_const => |val| {
try writer.writeByte(std.wasm.opcode(.f64_const));
try writer.writeInt(u64, @bitCast(val), .little);
},
.global_get => |val| {
try writer.writeByte(std.wasm.opcode(.global_get));
try leb.writeULEB128(writer, val);
},
}
try writer.writeByte(std.wasm.opcode(.end));
}
fn emitImport(wasm: *Wasm, writer: anytype, import: types.Import) !void {
const module_name = wasm.string_table.get(import.module_name);
try leb.writeULEB128(writer, @as(u32, @intCast(module_name.len)));
try writer.writeAll(module_name);
const name = wasm.string_table.get(import.name);
try leb.writeULEB128(writer, @as(u32, @intCast(name.len)));
try writer.writeAll(name);
try writer.writeByte(@intFromEnum(import.kind));
switch (import.kind) {
.function => |type_index| try leb.writeULEB128(writer, type_index),
.global => |global_type| {
try leb.writeULEB128(writer, std.wasm.valtype(global_type.valtype));
try writer.writeByte(@intFromBool(global_type.mutable));
},
.table => |table| {
try leb.writeULEB128(writer, std.wasm.reftype(table.reftype));
try emitLimits(writer, table.limits);
},
.memory => |limits| {
try emitLimits(writer, limits);
},
}
}
fn linkWithLLD(wasm: *Wasm, comp: *Compilation, prog_node: *std.Progress.Node) !void {
const tracy = trace(@src());
defer tracy.end();
var arena_allocator = std.heap.ArenaAllocator.init(wasm.base.allocator);
defer arena_allocator.deinit();
const arena = arena_allocator.allocator();
const directory = wasm.base.options.emit.?.directory; // Just an alias to make it shorter to type.
const full_out_path = try directory.join(arena, &[_][]const u8{wasm.base.options.emit.?.sub_path});
// If there is no Zig code to compile, then we should skip flushing the output file because it
// will not be part of the linker line anyway.
const module_obj_path: ?[]const u8 = if (wasm.base.options.module != null) blk: {
try wasm.flushModule(comp, prog_node);
if (fs.path.dirname(full_out_path)) |dirname| {
break :blk try fs.path.join(arena, &.{ dirname, wasm.base.intermediary_basename.? });
} else {
break :blk wasm.base.intermediary_basename.?;
}
} else null;
var sub_prog_node = prog_node.start("LLD Link", 0);
sub_prog_node.activate();
sub_prog_node.context.refresh();
defer sub_prog_node.end();
const is_obj = wasm.base.options.output_mode == .Obj;
const compiler_rt_path: ?[]const u8 = blk: {
if (comp.compiler_rt_lib) |lib| break :blk lib.full_object_path;
if (comp.compiler_rt_obj) |obj| break :blk obj.full_object_path;
break :blk null;
};
const target = wasm.base.options.target;
const id_symlink_basename = "lld.id";
var man: Cache.Manifest = undefined;
defer if (!wasm.base.options.disable_lld_caching) man.deinit();
var digest: [Cache.hex_digest_len]u8 = undefined;
if (!wasm.base.options.disable_lld_caching) {
man = comp.cache_parent.obtain();
// We are about to obtain this lock, so here we give other processes a chance first.
wasm.base.releaseLock();
comptime assert(Compilation.link_hash_implementation_version == 10);
for (wasm.base.options.objects) |obj| {
_ = try man.addFile(obj.path, null);
man.hash.add(obj.must_link);
}
for (comp.c_object_table.keys()) |key| {
_ = try man.addFile(key.status.success.object_path, null);
}
try man.addOptionalFile(module_obj_path);
try man.addOptionalFile(compiler_rt_path);
man.hash.addOptionalBytes(wasm.base.options.entry);
man.hash.addOptional(wasm.base.options.stack_size_override);
man.hash.add(wasm.base.options.build_id);
man.hash.add(wasm.base.options.import_memory);
man.hash.add(wasm.base.options.export_memory);
man.hash.add(wasm.base.options.import_table);
man.hash.add(wasm.base.options.export_table);
man.hash.addOptional(wasm.base.options.initial_memory);
man.hash.addOptional(wasm.base.options.max_memory);
man.hash.add(wasm.base.options.shared_memory);
man.hash.addOptional(wasm.base.options.global_base);
man.hash.add(wasm.base.options.export_symbol_names.len);
// strip does not need to go into the linker hash because it is part of the hash namespace
for (wasm.base.options.export_symbol_names) |symbol_name| {
man.hash.addBytes(symbol_name);
}
// We don't actually care whether it's a cache hit or miss; we just need the digest and the lock.
_ = try man.hit();
digest = man.final();
var prev_digest_buf: [digest.len]u8 = undefined;
const prev_digest: []u8 = Cache.readSmallFile(
directory.handle,
id_symlink_basename,
&prev_digest_buf,
) catch |err| blk: {
log.debug("WASM LLD new_digest={s} error: {s}", .{ std.fmt.fmtSliceHexLower(&digest), @errorName(err) });
// Handle this as a cache miss.
break :blk prev_digest_buf[0..0];
};
if (mem.eql(u8, prev_digest, &digest)) {
log.debug("WASM LLD digest={s} match - skipping invocation", .{std.fmt.fmtSliceHexLower(&digest)});
// Hot diggity dog! The output binary is already there.
wasm.base.lock = man.toOwnedLock();
return;
}
log.debug("WASM LLD prev_digest={s} new_digest={s}", .{ std.fmt.fmtSliceHexLower(prev_digest), std.fmt.fmtSliceHexLower(&digest) });
// We are about to change the output file to be different, so we invalidate the build hash now.
directory.handle.deleteFile(id_symlink_basename) catch |err| switch (err) {
error.FileNotFound => {},
else => |e| return e,
};
}
if (is_obj) {
// LLD's WASM driver does not support the equivalent of `-r` so we do a simple file copy
// here. TODO: think carefully about how we can avoid this redundant operation when doing
// build-obj. See also the corresponding TODO in linkAsArchive.
const the_object_path = blk: {
if (wasm.base.options.objects.len != 0)
break :blk wasm.base.options.objects[0].path;
if (comp.c_object_table.count() != 0)
break :blk comp.c_object_table.keys()[0].status.success.object_path;
if (module_obj_path) |p|
break :blk p;
// TODO I think this is unreachable. Audit this situation when solving the above TODO
// regarding eliding redundant object -> object transformations.
return error.NoObjectsToLink;
};
// This can happen when using --enable-cache and using the stage1 backend. In this case
// we can skip the file copy.
if (!mem.eql(u8, the_object_path, full_out_path)) {
try fs.cwd().copyFile(the_object_path, fs.cwd(), full_out_path, .{});
}
} else {
// Create an LLD command line and invoke it.
var argv = std.ArrayList([]const u8).init(wasm.base.allocator);
defer argv.deinit();
// We will invoke ourselves as a child process to gain access to LLD.
// This is necessary because LLD does not behave properly as a library -
// it calls exit() and does not reset all global data between invocations.
const linker_command = "wasm-ld";
try argv.appendSlice(&[_][]const u8{ comp.self_exe_path.?, linker_command });
try argv.append("--error-limit=0");
if (wasm.base.options.lto) {
switch (wasm.base.options.optimize_mode) {
.Debug => {},
.ReleaseSmall => try argv.append("-O2"),
.ReleaseFast, .ReleaseSafe => try argv.append("-O3"),
}
}
if (wasm.base.options.import_memory) {
try argv.append("--import-memory");
}
if (wasm.base.options.export_memory) {
try argv.append("--export-memory");
}
if (wasm.base.options.import_table) {
assert(!wasm.base.options.export_table);
try argv.append("--import-table");
}
if (wasm.base.options.export_table) {
assert(!wasm.base.options.import_table);
try argv.append("--export-table");
}
if (wasm.base.options.strip) {
try argv.append("-s");
}
if (wasm.base.options.initial_memory) |initial_memory| {
const arg = try std.fmt.allocPrint(arena, "--initial-memory={d}", .{initial_memory});
try argv.append(arg);
}
if (wasm.base.options.max_memory) |max_memory| {
const arg = try std.fmt.allocPrint(arena, "--max-memory={d}", .{max_memory});
try argv.append(arg);
}
if (wasm.base.options.shared_memory) {
try argv.append("--shared-memory");
}
if (wasm.base.options.global_base) |global_base| {
const arg = try std.fmt.allocPrint(arena, "--global-base={d}", .{global_base});
try argv.append(arg);
} else {
// We prepend it by default, so when a stack overflow happens the runtime will trap correctly,
// rather than silently overwrite all global declarations. See https://github.com/ziglang/zig/issues/4496
//
// The user can overwrite this behavior by setting the global-base
try argv.append("--stack-first");
}
// Users are allowed to specify which symbols they want to export to the wasm host.
for (wasm.base.options.export_symbol_names) |symbol_name| {
const arg = try std.fmt.allocPrint(arena, "--export={s}", .{symbol_name});
try argv.append(arg);
}
if (wasm.base.options.rdynamic) {
try argv.append("--export-dynamic");
}
if (wasm.base.options.entry) |entry| {
try argv.append("--entry");
try argv.append(entry);
}
// Increase the default stack size to a more reasonable value of 1MB instead of
// the default of 1 Wasm page being 64KB, unless overridden by the user.
try argv.append("-z");
const stack_size = wasm.base.options.stack_size_override orelse std.wasm.page_size * 16;
const arg = try std.fmt.allocPrint(arena, "stack-size={d}", .{stack_size});
try argv.append(arg);
if (wasm.base.options.output_mode == .Exe) {
if (wasm.base.options.wasi_exec_model == .reactor) {
// Reactor execution model does not have _start so lld doesn't look for it.
try argv.append("--no-entry");
// Make sure "_initialize" and other used-defined functions are exported if this is WASI reactor.
// If rdynamic is true, it will already be appended, so only verify if the user did not specify
// the flag in which case, we ensure `--export-dynamic` is called.
if (!wasm.base.options.rdynamic) {
try argv.append("--export-dynamic");
}
}
} else if (wasm.base.options.entry == null) {
try argv.append("--no-entry"); // So lld doesn't look for _start.
}
if (wasm.base.options.import_symbols) {
try argv.append("--allow-undefined");
}
// XXX - TODO: add when wasm-ld supports --build-id.
// if (wasm.base.options.build_id) {
// try argv.append("--build-id=tree");
// }
try argv.appendSlice(&.{ "-o", full_out_path });
if (target.cpu.arch == .wasm64) {
try argv.append("-mwasm64");
}
if (target.os.tag == .wasi) {
const is_exe_or_dyn_lib = wasm.base.options.output_mode == .Exe or
(wasm.base.options.output_mode == .Lib and wasm.base.options.link_mode == .Dynamic);
if (is_exe_or_dyn_lib) {
const wasi_emulated_libs = wasm.base.options.wasi_emulated_libs;
for (wasi_emulated_libs) |crt_file| {
try argv.append(try comp.get_libc_crt_file(
arena,
wasi_libc.emulatedLibCRFileLibName(crt_file),
));
}
if (wasm.base.options.link_libc) {
try argv.append(try comp.get_libc_crt_file(
arena,
wasi_libc.execModelCrtFileFullName(wasm.base.options.wasi_exec_model),
));
try argv.append(try comp.get_libc_crt_file(arena, "libc.a"));
}
if (wasm.base.options.link_libcpp) {
try argv.append(comp.libcxx_static_lib.?.full_object_path);
try argv.append(comp.libcxxabi_static_lib.?.full_object_path);
}
}
}
// Positional arguments to the linker such as object files.
var whole_archive = false;
for (wasm.base.options.objects) |obj| {
if (obj.must_link and !whole_archive) {
try argv.append("-whole-archive");
whole_archive = true;
} else if (!obj.must_link and whole_archive) {
try argv.append("-no-whole-archive");
whole_archive = false;
}
try argv.append(obj.path);
}
if (whole_archive) {
try argv.append("-no-whole-archive");
whole_archive = false;
}
for (comp.c_object_table.keys()) |key| {
try argv.append(key.status.success.object_path);
}
if (module_obj_path) |p| {
try argv.append(p);
}
if (wasm.base.options.output_mode != .Obj and
!wasm.base.options.skip_linker_dependencies and
!wasm.base.options.link_libc)
{
try argv.append(comp.libc_static_lib.?.full_object_path);
}
if (compiler_rt_path) |p| {
try argv.append(p);
}
if (wasm.base.options.verbose_link) {
// Skip over our own name so that the LLD linker name is the first argv item.
Compilation.dump_argv(argv.items[1..]);
}
if (std.process.can_spawn) {
// If possible, we run LLD as a child process because it does not always
// behave properly as a library, unfortunately.
// https://github.com/ziglang/zig/issues/3825
var child = std.ChildProcess.init(argv.items, arena);
if (comp.clang_passthrough_mode) {
child.stdin_behavior = .Inherit;
child.stdout_behavior = .Inherit;
child.stderr_behavior = .Inherit;
const term = child.spawnAndWait() catch |err| {
log.err("unable to spawn {s}: {s}", .{ argv.items[0], @errorName(err) });
return error.UnableToSpawnWasm;
};
switch (term) {
.Exited => |code| {
if (code != 0) {
std.process.exit(code);
}
},
else => std.process.abort(),
}
} else {
child.stdin_behavior = .Ignore;
child.stdout_behavior = .Ignore;
child.stderr_behavior = .Pipe;
try child.spawn();
const stderr = try child.stderr.?.reader().readAllAlloc(arena, std.math.maxInt(usize));
const term = child.wait() catch |err| {
log.err("unable to spawn {s}: {s}", .{ argv.items[0], @errorName(err) });
return error.UnableToSpawnWasm;
};
switch (term) {
.Exited => |code| {
if (code != 0) {
comp.lockAndParseLldStderr(linker_command, stderr);
return error.LLDReportedFailure;
}
},
else => {
log.err("{s} terminated with stderr:\n{s}", .{ argv.items[0], stderr });
return error.LLDCrashed;
},
}
if (stderr.len != 0) {
log.warn("unexpected LLD stderr:\n{s}", .{stderr});
}
}
} else {
const exit_code = try lldMain(arena, argv.items, false);
if (exit_code != 0) {
if (comp.clang_passthrough_mode) {
std.process.exit(exit_code);
} else {
return error.LLDReportedFailure;
}
}
}
// Give +x to the .wasm file if it is an executable and the OS is WASI.
// Some systems may be configured to execute such binaries directly. Even if that
// is not the case, it means we will get "exec format error" when trying to run
// it, and then can react to that in the same way as trying to run an ELF file
// from a foreign CPU architecture.
if (fs.has_executable_bit and target.os.tag == .wasi and
wasm.base.options.output_mode == .Exe)
{
// TODO: what's our strategy for reporting linker errors from this function?
// report a nice error here with the file path if it fails instead of
// just returning the error code.
// chmod does not interact with umask, so we use a conservative -rwxr--r-- here.
try std.os.fchmodat(fs.cwd().fd, full_out_path, 0o744, 0);
}
}
if (!wasm.base.options.disable_lld_caching) {
// Update the file with the digest. If it fails we can continue; it only
// means that the next invocation will have an unnecessary cache miss.
Cache.writeSmallFile(directory.handle, id_symlink_basename, &digest) catch |err| {
log.warn("failed to save linking hash digest symlink: {s}", .{@errorName(err)});
};
// Again failure here only means an unnecessary cache miss.
man.writeManifest() catch |err| {
log.warn("failed to write cache manifest when linking: {s}", .{@errorName(err)});
};
// We hang on to this lock so that the output file path can be used without
// other processes clobbering it.
wasm.base.lock = man.toOwnedLock();
}
}
fn reserveVecSectionHeader(bytes: *std.ArrayList(u8)) !u32 {
// section id + fixed leb contents size + fixed leb vector length
const header_size = 1 + 5 + 5;
const offset = @as(u32, @intCast(bytes.items.len));
try bytes.appendSlice(&[_]u8{0} ** header_size);
return offset;
}
fn reserveCustomSectionHeader(bytes: *std.ArrayList(u8)) !u32 {
// unlike regular section, we don't emit the count
const header_size = 1 + 5;
const offset = @as(u32, @intCast(bytes.items.len));
try bytes.appendSlice(&[_]u8{0} ** header_size);
return offset;
}
fn writeVecSectionHeader(buffer: []u8, offset: u32, section: std.wasm.Section, size: u32, items: u32) !void {
var buf: [1 + 5 + 5]u8 = undefined;
buf[0] = @intFromEnum(section);
leb.writeUnsignedFixed(5, buf[1..6], size);
leb.writeUnsignedFixed(5, buf[6..], items);
buffer[offset..][0..buf.len].* = buf;
}
fn writeCustomSectionHeader(buffer: []u8, offset: u32, size: u32) !void {
var buf: [1 + 5]u8 = undefined;
buf[0] = 0; // 0 = 'custom' section
leb.writeUnsignedFixed(5, buf[1..6], size);
buffer[offset..][0..buf.len].* = buf;
}
fn emitLinkSection(wasm: *Wasm, binary_bytes: *std.ArrayList(u8), symbol_table: *std.AutoArrayHashMap(SymbolLoc, u32)) !void {
const offset = try reserveCustomSectionHeader(binary_bytes);
const writer = binary_bytes.writer();
// emit "linking" custom section name
const section_name = "linking";
try leb.writeULEB128(writer, section_name.len);
try writer.writeAll(section_name);
// meta data version, which is currently '2'
try leb.writeULEB128(writer, @as(u32, 2));
// For each subsection type (found in types.Subsection) we can emit a section.
// Currently, we only support emitting segment info and the symbol table.
try wasm.emitSymbolTable(binary_bytes, symbol_table);
try wasm.emitSegmentInfo(binary_bytes);
const size = @as(u32, @intCast(binary_bytes.items.len - offset - 6));
try writeCustomSectionHeader(binary_bytes.items, offset, size);
}
fn emitSymbolTable(wasm: *Wasm, binary_bytes: *std.ArrayList(u8), symbol_table: *std.AutoArrayHashMap(SymbolLoc, u32)) !void {
const writer = binary_bytes.writer();
try leb.writeULEB128(writer, @intFromEnum(types.SubsectionType.WASM_SYMBOL_TABLE));
const table_offset = binary_bytes.items.len;
var symbol_count: u32 = 0;
for (wasm.resolved_symbols.keys()) |sym_loc| {
const symbol = sym_loc.getSymbol(wasm).*;
if (symbol.tag == .dead) continue; // Do not emit dead symbols
try symbol_table.putNoClobber(sym_loc, symbol_count);
symbol_count += 1;
log.debug("Emit symbol: {}", .{symbol});
try leb.writeULEB128(writer, @intFromEnum(symbol.tag));
try leb.writeULEB128(writer, symbol.flags);
const sym_name = if (wasm.export_names.get(sym_loc)) |exp_name| wasm.string_table.get(exp_name) else sym_loc.getName(wasm);
switch (symbol.tag) {
.data => {
try leb.writeULEB128(writer, @as(u32, @intCast(sym_name.len)));
try writer.writeAll(sym_name);
if (symbol.isDefined()) {
try leb.writeULEB128(writer, symbol.index);
const atom_index = wasm.symbol_atom.get(sym_loc).?;
const atom = wasm.getAtom(atom_index);
try leb.writeULEB128(writer, @as(u32, atom.offset));
try leb.writeULEB128(writer, @as(u32, atom.size));
}
},
.section => {
try leb.writeULEB128(writer, symbol.index);
},
else => {
try leb.writeULEB128(writer, symbol.index);
if (symbol.isDefined()) {
try leb.writeULEB128(writer, @as(u32, @intCast(sym_name.len)));
try writer.writeAll(sym_name);
}
},
}
}
var buf: [10]u8 = undefined;
leb.writeUnsignedFixed(5, buf[0..5], @as(u32, @intCast(binary_bytes.items.len - table_offset + 5)));
leb.writeUnsignedFixed(5, buf[5..], symbol_count);
try binary_bytes.insertSlice(table_offset, &buf);
}
fn emitSegmentInfo(wasm: *Wasm, binary_bytes: *std.ArrayList(u8)) !void {
const writer = binary_bytes.writer();
try leb.writeULEB128(writer, @intFromEnum(types.SubsectionType.WASM_SEGMENT_INFO));
const segment_offset = binary_bytes.items.len;
try leb.writeULEB128(writer, @as(u32, @intCast(wasm.segment_info.count())));
for (wasm.segment_info.values()) |segment_info| {
log.debug("Emit segment: {s} align({d}) flags({b})", .{
segment_info.name,
segment_info.alignment,
segment_info.flags,
});
try leb.writeULEB128(writer, @as(u32, @intCast(segment_info.name.len)));
try writer.writeAll(segment_info.name);
try leb.writeULEB128(writer, segment_info.alignment.toLog2Units());
try leb.writeULEB128(writer, segment_info.flags);
}
var buf: [5]u8 = undefined;
leb.writeUnsignedFixed(5, &buf, @as(u32, @intCast(binary_bytes.items.len - segment_offset)));
try binary_bytes.insertSlice(segment_offset, &buf);
}
pub fn getULEB128Size(uint_value: anytype) u32 {
const T = @TypeOf(uint_value);
const U = if (@typeInfo(T).Int.bits < 8) u8 else T;
var value = @as(U, @intCast(uint_value));
var size: u32 = 0;
while (value != 0) : (size += 1) {
value >>= 7;
}
return size;
}
/// For each relocatable section, emits a custom "relocation.<section_name>" section
fn emitCodeRelocations(
wasm: *Wasm,
binary_bytes: *std.ArrayList(u8),
section_index: u32,
symbol_table: std.AutoArrayHashMap(SymbolLoc, u32),
) !void {
const code_index = wasm.code_section_index orelse return;
const writer = binary_bytes.writer();
const header_offset = try reserveCustomSectionHeader(binary_bytes);
// write custom section information
const name = "reloc.CODE";
try leb.writeULEB128(writer, @as(u32, @intCast(name.len)));
try writer.writeAll(name);
try leb.writeULEB128(writer, section_index);
const reloc_start = binary_bytes.items.len;
var count: u32 = 0;
var atom: *Atom = wasm.getAtomPtr(wasm.atoms.get(code_index).?);
// for each atom, we calculate the uleb size and append that
var size_offset: u32 = 5; // account for code section size leb128
while (true) {
size_offset += getULEB128Size(atom.size);
for (atom.relocs.items) |relocation| {
count += 1;
const sym_loc: SymbolLoc = .{ .file = atom.file, .index = relocation.index };
const symbol_index = symbol_table.get(sym_loc).?;
try leb.writeULEB128(writer, @intFromEnum(relocation.relocation_type));
const offset = atom.offset + relocation.offset + size_offset;
try leb.writeULEB128(writer, offset);
try leb.writeULEB128(writer, symbol_index);
if (relocation.relocation_type.addendIsPresent()) {
try leb.writeILEB128(writer, relocation.addend);
}
log.debug("Emit relocation: {}", .{relocation});
}
atom = if (atom.prev) |prev| wasm.getAtomPtr(prev) else break;
}
if (count == 0) return;
var buf: [5]u8 = undefined;
leb.writeUnsignedFixed(5, &buf, count);
try binary_bytes.insertSlice(reloc_start, &buf);
const size = @as(u32, @intCast(binary_bytes.items.len - header_offset - 6));
try writeCustomSectionHeader(binary_bytes.items, header_offset, size);
}
fn emitDataRelocations(
wasm: *Wasm,
binary_bytes: *std.ArrayList(u8),
section_index: u32,
symbol_table: std.AutoArrayHashMap(SymbolLoc, u32),
) !void {
if (wasm.data_segments.count() == 0) return;
const writer = binary_bytes.writer();
const header_offset = try reserveCustomSectionHeader(binary_bytes);
// write custom section information
const name = "reloc.DATA";
try leb.writeULEB128(writer, @as(u32, @intCast(name.len)));
try writer.writeAll(name);
try leb.writeULEB128(writer, section_index);
const reloc_start = binary_bytes.items.len;
var count: u32 = 0;
// for each atom, we calculate the uleb size and append that
var size_offset: u32 = 5; // account for code section size leb128
for (wasm.data_segments.values()) |segment_index| {
var atom: *Atom = wasm.getAtomPtr(wasm.atoms.get(segment_index).?);
while (true) {
size_offset += getULEB128Size(atom.size);
for (atom.relocs.items) |relocation| {
count += 1;
const sym_loc: SymbolLoc = .{
.file = atom.file,
.index = relocation.index,
};
const symbol_index = symbol_table.get(sym_loc).?;
try leb.writeULEB128(writer, @intFromEnum(relocation.relocation_type));
const offset = atom.offset + relocation.offset + size_offset;
try leb.writeULEB128(writer, offset);
try leb.writeULEB128(writer, symbol_index);
if (relocation.relocation_type.addendIsPresent()) {
try leb.writeILEB128(writer, relocation.addend);
}
log.debug("Emit relocation: {}", .{relocation});
}
atom = if (atom.prev) |prev| wasm.getAtomPtr(prev) else break;
}
}
if (count == 0) return;
var buf: [5]u8 = undefined;
leb.writeUnsignedFixed(5, &buf, count);
try binary_bytes.insertSlice(reloc_start, &buf);
const size = @as(u32, @intCast(binary_bytes.items.len - header_offset - 6));
try writeCustomSectionHeader(binary_bytes.items, header_offset, size);
}
fn hasPassiveInitializationSegments(wasm: *const Wasm) bool {
var it = wasm.data_segments.iterator();
while (it.next()) |entry| {
const segment: Segment = wasm.segments.items[entry.value_ptr.*];
if (segment.needsPassiveInitialization(wasm.base.options.import_memory, entry.key_ptr.*)) {
return true;
}
}
return false;
}
pub fn getTypeIndex(wasm: *const Wasm, func_type: std.wasm.Type) ?u32 {
var index: u32 = 0;
while (index < wasm.func_types.items.len) : (index += 1) {
if (wasm.func_types.items[index].eql(func_type)) return index;
}
return null;
}
/// Searches for a matching function signature. When no matching signature is found,
/// a new entry will be made. The value returned is the index of the type within `wasm.func_types`.
pub fn putOrGetFuncType(wasm: *Wasm, func_type: std.wasm.Type) !u32 {
if (wasm.getTypeIndex(func_type)) |index| {
return index;
}
// functype does not exist.
const index = @as(u32, @intCast(wasm.func_types.items.len));
const params = try wasm.base.allocator.dupe(std.wasm.Valtype, func_type.params);
errdefer wasm.base.allocator.free(params);
const returns = try wasm.base.allocator.dupe(std.wasm.Valtype, func_type.returns);
errdefer wasm.base.allocator.free(returns);
try wasm.func_types.append(wasm.base.allocator, .{
.params = params,
.returns = returns,
});
return index;
}
/// For the given `decl_index`, stores the corresponding type representing the function signature.
/// Asserts declaration has an associated `Atom`.
/// Returns the index into the list of types.
pub fn storeDeclType(wasm: *Wasm, decl_index: Module.Decl.Index, func_type: std.wasm.Type) !u32 {
const atom_index = wasm.decls.get(decl_index).?;
const index = try wasm.putOrGetFuncType(func_type);
try wasm.atom_types.put(wasm.base.allocator, atom_index, index);
return index;
}